wiggleverse/session-history
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
e4ec14379c7cca2111002b3bd5bb29611e4560d4
session-history/0028/SESSION-0028.0-TRANSCRIPT-2026-05-28T14-05--2026-05-28T14-32.md
T

12 KiB
Raw Blame History

Session 0028.0 — Transcript

Date: 2026-05-28 Start: 2026-05-28T14-05 (PST implied) End: 2026-05-28T14-32 (PST implied) Subject: flotilla v1.2.0 — deploy robustness + record-maintenance verbs (PR, not merged) Status: FINALIZED.

Plan

Item picked: flotilla v1.2.0 — deploy-robustness + record-maintenance verbs (the §19.2 infra candidate surfaced by Session 0022). Pure-flotilla, touches no rfc-app code → zero collision with the in-flight rfc-app feature sessions 0026/0027. No OHM deploy, no version-slot contention, no secret bytes.

Working on branch session-0028/flotilla-v1.2.0; shipping as a PR (operator merges + tags — per the "let's use branches and PRs" directive this session).

Deliverables:

  1. Preflight SSH-failure disambiguation (deploy.py phase 2). Today any non-zero test -d <install> is reported "install dir missing" — but a gcloud compute ssh connection/auth/IAP failure also returns non-zero, so a connectivity failure is misreported as a missing dir (deploy.py:355). Fix: a connectivity probe runs first; its failure reports the real cause.
  2. SshError → clean phase failure. _PhaseRunner.run only catches _PhaseFailure; an ssh.SshError (e.g. a restart SSH timeout) propagates and leaves the deploys row stuck in_progress holding the §8.3 lock. Catch it, record a phase failure, finalize the row.
  3. SSE pre-drain restart (phase 7). Replace systemctl restart with a flotilla-controlled stop (SIGTERM → bounded drain window → SIGKILL fallback for lingering SSE keepalive streams) → start. Bounds the window flotilla waits regardless of the unit's TimeoutStopSec. This is the v0.6.0 (Session 0009/"I") failure made un-reproducible from the flotilla side. A graceful app-cooperative drain endpoint in rfc-app stays a future §19.2 candidate.
  4. flotilla deployment update verb (registry + CLI) — update mutable record fields without remove+add. Motivated by the #1 VM rename.
  5. flotilla deploy reconcile verb — probe /api/health and reconcile the deploy log to reality: finalize a stuck in_progress row as succeeded when health confirms the target version (releases the lock); honestly report terminal rows health contradicts without falsifying the audit trail.

Each lands with unit tests (no network — the deploy seams are DI'd). CHANGELOG 1.2.0 + VERSION/pyproject bump. Transcript finalized + published LAST.

Parallel context at claim: sessions 0026 + 0027 also in-progress (launch prompts not captured in their placeholders). flotilla-only scope chosen specifically to stay disjoint from whatever rfc-app surfaces they touch.

Friction log

  • 14:09 — shared-working-directory collision. Branches alone don't isolate concurrent sessions: git tracks one working tree per checkout. Shortly after I branched in the canonical flotilla dir, another live session (0026 or 0027) wrote ohm_rfc_app_flotilla/deploy.py + tests/test_deploy.py in the same directory (a secret-redaction hardening, "finding L1"). Their files overlap exactly with mine (_PhaseRunner, phase details). Editing in place would corrupt both sessions' work. Resolution: restored the shared dir to main with their unstaged edits preserved, then created a manual git worktree at …/ohm-rfc-app-flotilla-s0028 on my branch with its own fresh venv (pip install -e .[test]; the canonical .venv's PEP 660 editable finder is pinned to the shared dir, so a shared venv would defeat isolation). Worktree baseline: 157 tests green on clean c26f251. Handoff implication: parallel flotilla sessions MUST each take a worktree; the operator's "branches and PRs" guidance needs the worktree step to actually isolate. Their redaction change and my v1.2.0 change both touch deploy.py → expect a merge conflict in _PhaseRunner when the second PR merges; resolvable (different concerns: they redact phase detail, I add an SshError branch + DI redactor coexist cleanly).

Launch prompt

You are an OHM roadmap driver session, running in parallel with others. Claim session ID first. STATE (end of 0022): OHM live at rfc-app v0.24.0 on the renamed VM ohm-rfc-app; deploys unblocked. Pick ONE disjoint roadmap item. Strong infra candidate: flotilla v1.2.0 — fix the preflight that misreports an SSH/auth failure as "install dir missing" (deploy.py:355), add a deployment update verb, add flotilla deploy reconcile + SSE pre-drain (§19.2 from this session). "Let's use branches and PR's so we don't need to worry about conflicting with other sessions."

What shipped

flotilla v1.2.0, on branch session-0028/flotilla-v1.2.0, pushed to origin (gitea). Open as a PR (not yet merged):

https://git.wiggleverse.org/wiggleverse/ohm-rfc-app-flotilla/pulls/new/session-0028/flotilla-v1.2.0

Commit 51c7ea8. No OHM deploy, no rfc-app release, no pin bump — this is operator-tool code. Operator opens + merges the PR (sole-merger rule) and tags v1.2.0 after merge.

Five changes, all flotilla-internal, each unit-tested (the deploy seams are dependency-injected so nothing touches gcloud or the network):

  1. Preflight SSH-failure disambiguation (deploy.py phase 2). The misreport at deploy.py:355: a bare test -d <install> whose any-non-zero return was read as "install dir missing" — but gcloud compute ssh exits non-zero on a connection/auth/IAP failure too, so an unreachable VM was misdiagnosed as a missing dir. Fix: a trivial true connectivity probe runs first; its non-zero return is reported as the auth/connectivity failure it is. (Initially gold-plated with a stdout marker; dropped that for returncode-only after it would have forced every test stub — and every future one — to simulate the echo. Returncode-only fully solves the actual bug.)

  2. _PhaseRunner catches ssh.SshError → records a clean failed phase instead of unwinding past run_deploy and stranding the row in_progress holding the §8.3 lock. This is the structural cause of the v0.6.0 / Session 0009 "aborted but actually fine" — a restart SSH timeout raised mid-phase.

  3. Phase-7 SSE pre-drain. Replaced systemctl restart with a flotilla-controlled systemctl kill -s TERM → bounded drain_wait_seconds (default 20s) → systemctl kill -s KILL for lingering SSE keepalive streams → systemctl start. Bounded by drain + restart_wait regardless of the unit's TimeoutStopSec, so a held-open SSE stream can't blow past the SSH timeout. The phase detail records whether it drained via graceful SIGTERM or had to SIGKILL.

  4. flotilla deployment update <name> [--flag …] (registry + CLI). In-place edit of mutable record fields (VM name/zone/project/service user/install dir/systemd unit/tunnel flag, pin source fields, corpus repo, health URL) without the remove+add that drops the overlay, secret bindings, and deploy log. The verb the #1 VM rename wanted (session 0022 had to edit the record by hand).

  5. flotilla deploy reconcile <name> [--json]. Probes /api/health once and settles the log against reality: a stuck in_progress row is finalized succeeded when health confirms the target version (releasing the lock, without re-running the deploy); an already-terminal row is never rewritten — reconcile records a fresh health snapshot and reports the truth (including the "recorded failed but health says target is up" case the pre-drain fix makes rare). Exits non-zero when not confirmed healthy.

SPEC folded back (§19.3 rule 2): §8.1 step 7 rewritten for the pre-drain, §12.1 verb inventory gains the two verbs, and a new §19.2 candidate ("graceful app-cooperative SSE drain" — a framework-side drain endpoint in rfc-app that would make the SIGKILL fallback go unused) records the remaining future work.

Tests: 180 green (157 baseline + 23 new) in the isolated worktree venv.

Decisions

  • Scope = deploy-robustness, not GCP-resource alignment. flotilla v1.2.0 is referenced in roadmap #34 as a possible slot, but #34 is the GCP resource-rename arc. This is a different, disjoint scope; #34 is untouched and unclaimed. Did NOT edit ohm-rfc/ROADMAP.md (no rfc-app/OHM surface moved, and 0026/0027 may be editing it).
  • Pre-drain is flotilla-only. A graceful SIGTERM exit needs app cooperation (rfc-app closing SSE streams on a signal); that's cross-repo and stays a §19.2 candidate. The SIGKILL fallback is the flotilla-only mitigation that needs no framework change and makes the v0.6.0 fault un-reproducible from flotilla's side today.
  • reconcile never rewrites terminal rows. Finalizing a stuck in_progress is honest (the deploy did complete; only the terminal write was lost). Rewriting a failed/aborted row to succeeded would erase the audit trail, so those get a health snapshot + an advisory instead.
  • Branches+PR, no direct main push, no tag. Per the operator directive and the sole-merger rule, I pushed the branch and left the merge + tag to the operator.

Friction (continued)

  • No gitea PR-create automation available. tea CLI not installed; no gitea API token in env (HTTPS creds live in macOS keychain — not fished out). So the PR is "pushed + create-URL handed to operator," not auto-opened. The operator opens + merges anyway.

Operating notes / cleanup

  • Isolated worktree left in place at /Users/benstull/projects/wiggleverse/ohm-rfc-app-flotilla-s0028 (its own .venv) in case the operator wants to iterate on the PR. Remove with git worktree remove ohm-rfc-app-flotilla-s0028 after the PR merges.
  • The canonical dir …/ohm-rfc-app-flotilla was restored to main carrying the OTHER session's (0026/0027) unstaged secret-redaction edits, untouched — that work is theirs to commit on their branch.

Handoff prompt for the next session

(Delivered to the operator in chat before this transcript was published.)

You are an OHM roadmap driver session, running in parallel with others. CLAIM YOUR SESSION ID FIRST: ~/git/ohm-infra/scripts/claim-session-id.sh --start <now PST, YYYY-MM-DDTHH-MM> It prints the active in-progress sessions — coordinate if your work overlaps. Read ~/git/ohm-infra/SESSION-PROTOCOL.md §7 + PULL ohm-rfc/ROADMAP.md once.

STATE (end of 0028, 2026-05-28): OHM live at rfc-app v0.24.0 on the renamed VM ohm-rfc-app; deploys unblocked, lock free. flotilla v1.2.0 is built

CRITICAL for parallel flotilla work: branches alone do NOT isolate concurrent sessions sharing one working dir — take a manual git worktree (git worktree add <path> -b <branch> main) with its own venv (python3 -m venv .venv && .venv/bin/pip install -e .[test]; the canonical .venv's editable finder is pinned to the shared dir).

PICK ONE disjoint item. Open candidates: #28 (PR cross-references — large, its own session); #21 Part A Amplitude audit (~week of real data, ~2026-06-04); #22 pro-consent copy (needs operator-drafted + counsel text); #20/#18 ops (DMARC Phase B ~2026-06-04, stale wiggleverse/meta webhook still pointing at deprovisioned rfc.wiggleverse.org, bounce wiring); #31b/#25-redesign (need operator screenshots); #33/#34 repo + GCP name alignment (operator-led). AVOID shipped surfaces (#26 ProposeModal/PR/RFC views + proposed_use_cases; #27 suggest-tags; #29 sign-in resume + /api/auth/me + lib/useLastState.js; v0.21.0 polish surfaces).

Hard rules: never ask for secret bytes (give the pbpaste | flotilla secret set ohm-rfc-app <KEY> gesture); coordinate the serialized deploy lock + version slot with the operator; subagents write their own §5 subsession transcripts; finalize + publish your transcript LAST (deliver the handoff prompt before publish). Use branches + PRs. Claim your ID, then go.

Reference in New Issue View Git Blame Copy Permalink