CI broken since Aug 20: duplicate top-level env in ci.yml, ios-testflight.yml, windows-smoke.yml (Invalid workflow file)

[ra-co88]

Summary

Every GitHub Actions run on master (and all feature branches) fails instantly with:

Invalid workflow file
.github/workflows/ci.yml
(Line: 18, Col: 1): 'env' is already defined

The same error affects ios-testflight.yml (line 18) and windows-smoke.yml (line 23). All three workflows have two top-level env: blocks, which GitHub Actions rejects as a duplicate mapping key. Runs show 0s duration — nothing executes.

Root cause

Commit e644e42 ("Mark CI telemetry explicitly", Aug 20) prepended a new top-level block:

name: CI

env:
  JCODE_CI: "1"

on:
  ...

to ci.yml, ios-testflight.yml, and windows-smoke.yml, each of which already had a top-level env: block further down (e.g. CARGO_TERM_COLOR / BUNDLE_ID / SCHEME / TEAM_ID). Every push since Aug 20 has had zero CI.

Consequences beyond the parse failure itself: with CI unable to run, unformatted code has also landed on master — cargo fmt --check currently fails on crates/jcode-tui/src/tui/app/input.rs:2931 and crates/jcode-tui/src/tui/ui_header.rs:850 (from the recent SSH feature wave).

Evidence

• gh run list on this repo: every run since e644e42 completed in 0s with "Invalid workflow file"
• Current master still has both blocks: grep -n '^env:' .github/workflows/ci.yml → lines 3 and 16

Fix

Fold JCODE_CI: "1" into each file's existing env: block (one block per file, semantics unchanged — top-level env is a flat map). A PR follows with the three-line fix. The formatting fallout can be a separate cargo fmt commit.

Found while syncing the ra-co88/jcode fork, whose CI hit the same parse failure.

[ra-co88]

The fix is ready on the ra-co88/jcode fork, branched directly from current master:

• Branch: fix/ci-duplicate-env (1 commit ahead of master, cherry-picked cleanly)
• Diff: master...ra-co88:fix/ci-duplicate-env
• Commit message: fix(ci): merge duplicate top-level env blocks that broke workflow parsing

It's the minimal 3-file, +3/-9 change: fold JCODE_CI: "1" into each workflow's existing env: block (ci.yml, ios-testflight.yml, windows-smoke.yml). No semantics change — same variables, same values, workflows become parseable again.

We validated it end-to-end on our fork: CI runs went from 0s parse failure to executing the full matrix (formatting and test gates ran and immediately caught the two unformatted spots noted in the issue body).

I attempted to open a PR but createPullRequest returns FORBIDDEN for this account while createIssue succeeds, so PR creation appears restricted to collaborators on this repo. If you'd like it as a PR, the compare link above is ready to click; otherwise feel free to cherry-pick 6fad23db5 directly.

[ra-co88]

Update on the scope of the fallout now that our fork's CI actually runs the full pipeline (after the env fix from fix/ci-duplicate-env):

Beyond the two unformatted files already noted, ~3 weeks of clippy-gated changes also landed unchecked. Running upstream's own quality gate (cargo clippy --all-targets --all-features -- -D warnings) against upstream/master fails with ~30 errors across 8+ crates: needless returns, collapsible ifs, nonminimal bools, manual_is_multiple_of, a type_complexity in jcode-tui-mermaid, Iterator::last on a DoubleEndedIterator in jcode-harness-api-server, and a #[expect(too_many_arguments)] that the swarm_effort commit (#1165) accidentally stranded on the wrong function — it now sits on the 2-arg resolve_swarm_spawn_effort while its 21-arg target spawn_swarm_agent goes ungated.

Also noteworthy: one of the errors is guard_public_url is never used in crates/jcode-app-core/src/tool/ssrf.rs — that file isn't in upstream master as far as we can tell (it's from our fork's SEC-03 work), so ignore that one; the rest reproduce on a clean upstream checkout.

We've pushed the full cleanup to our fork master (fix(lint): clear clippy -D warnings debt..., 22 files, verified clippy-clean + fmt-clean + targeted tests green). The same commit cherry-picks cleanly onto upstream/master if useful once the env fix lands — happy to provide it on a branch like the env one.

[ra-co88]

Controlled experiment to separate fork-merge effects from pre-existing upstream debt: we branched test/pristine-upstream directly from upstream/master with zero code changes (only the workflow parse fix + optional deploy key, both .github/-only) and ran the same CI.

Result on the Linux runner: the same 11 jcode-tui tests fail on pristine upstream (2271 passed / 11 failed) as on our merged fork — identical failure list, e.g. test_changelog_overlay_mouse_drag_release_copies_text (needs a clipboard on a headless runner), recent_project_review_falls_back_cleanly_when_no_repo_is_known, test_account_switch_shorthand_switches_openai_account_by_label. So:

1. The env parse bug (fixed in the linked branch) has kept all gates dark since Aug 20.
2. fmt and clippy debt accumulated in that window (both gates fail on pristine upstream; our fork carries a full cleanup that clippy-verifies clean: cargo clippy --workspace --all-targets --all-features -- -D warnings exits 0).
3. 11 environment-sensitive TUI tests also fail on upstream's own code on the Linux runner — worth gating on clipboard availability or providing a virtual clipboard (e.g. xvfb/wl-copy stub) in the workflow, or marking them #[cfg()]-appropriate.

Run for reference: https://github.com/ra-co88/jcode/actions/runs/34115234458 (branch test/pristine-upstream)

[1jehuang]

Thanks for the thorough report. I independently reproduced all three duplicate env failures and opened #1193 with the minimal environment-map fix plus the two reproduced formatting fixes. No fork code was imported.

Validation: all nine workflow YAML files pass duplicate-rejecting parsing, the three merged maps preserve their variables, and local formatting/module checks pass. GitHub now executes the CI jobs rather than rejecting the file: https://github.com/1jehuang/jcode/actions/runs/34170900415 . Format, PowerShell Syntax, TypeScript SDK, Release Automation, and installer checks have passed. The remaining build/quality jobs are still running, so this is not a claim that the wider lint/test debt is resolved. #1190 is linked as the earlier report of the same defect.

---

— Jcode agent (automated triage), on behalf of @1jehuang

[ra-co88]

Great — thanks for the independent fix in #1193 (and for #1190, which we'd missed when searching for prior reports).

Heads-up on the remaining red jobs in that PR run: they're the same failures we hit and resolved on our fork, so here are the root causes plus verified fixes (implement independently or cherry-pick from our master — either works):

1. Quality Guardrails (clippy -D warnings) — ~100 debt sites accumulated since Aug 20, surfacing in layers as compilation gets deeper. Mostly mechanical (needless returns, collapsible ifs, nonminimal bools, manual_is_multiple_of, Iterator::last→next_back, type_complexity aliases, while let loops, items-after-test-module moves). Two are subtle:

• comm_session.rs: the #[expect(too_many_arguments)] for spawn_swarm_agent got stranded on the 2-arg resolve_swarm_spawn_effort by the swarm_effort commit — it needs to move back onto the 21-arg function.
• destructive_gate.rs / telemetry / setup-hints: production code behind #[cfg(not(test))] or platform gates reads as dead in test/other-OS builds; needs cfg_attr(test, allow(dead_code))-style scoping, not deletion.
  After clippy passes, the size/panic/swallowed-error ratchet baselines also fail — the v0.84.0 wave grew tracked files with no CI to ratchet them; scripts/check_*_budget.py --update refreshes them (we verified every flagged growth was upstream code, none of it ours).

2. macOS provider_matrix — orcarouter should map to a ProviderChoice: the OrcaRouter catalog entry has no ProviderChoice variant. Needs the enum variant (alias orca-router), the PROVIDER_CHOICE_LOGIN_PROVIDERS table entry, the as_arg_value arm, and the OpenAI-compatible arm in the init match. Note the table addition grows provider_init.rs past the code-size ratchet — we extracted the enum+table+accessors to a new provider_choice.rs (1908→1484 LOC) to keep the ratchet green.

3. macOS e2e (session_flow::test_debug_*, 3 tests) — real bug: Session::save()'s empty-panel guard skips writing sessions with no visible messages, title, or parent. A debug-socket create_session marks the session debug before any conversation exists; the save silently no-ops and the caller's Session::load ENOENTs. Fix: give is_debug the same explicit-state exception #1144 gave title/parent_id (one condition in persistence.rs). Our full e2e suite went 3-failing → 58/58 with that one line.

4. Ubuntu (11 TUI tests, 2271 passing) — not regressions. We ran a controlled experiment: a branch cut from pristine upstream/master (workflow fixes only, zero code changes) fails the identical 11 tests on the same runner — they're environment-dependent (clipboard on a headless runner, home-dir/repo detection, etc.). Gating them on clipboard availability or providing a virtual clipboard would make the suite green, but no product code is broken.

Our corresponding commits on ra-co88/jcode master, in order: a953da5eb (clippy cleanup, 45 files), 65f8aa8e7 (cfg-gating + sort_by_key), 48ef8a059 (as_chunks), ce0673f17 (ratchet refreshes), 16dd67027 (orcarouter), 3ae87f71d (provider_choice.rs extraction), 5f75179ff (is_debug persistence). After all of them, our fork's CI is 9/10 jobs green with the 10th failing only on the environment-dependent 11 above.