[huong8373tt-beep]

Summary

Treat ~/.teamai/config.yaml without a repo key as a settings-only machine file rather than an initialized user installation. When init --scope user promotes that file, preserve valid member settings and skip invalid values with a warning.

This is step 1 of #1459, so this PR references the issue instead of closing it.

Validation

• npm run build — passed
• npm run typecheck — passed
• npm run lint — passed
• npm test -- src/tests/config-not-initialized.test.ts — 23 passed
• git diff --check — passed
• CI at commit 3ca7b95 is green; the cancelled fork-safe E2E entry belongs to an earlier cancelled run.

Built CLI verification recorded in review

The maintainer ran the built CLI with an isolated HOME and a local bare Git remote. The pre-seeded settings-only ~/.teamai/config.yaml was:

recallEnabled: "yes"
coAuthorEnabled: false
ai:
  tool: codex
workflows:
  tools:
    default: codex

Commands exercised:

teamai status
teamai pull
teamai init <synthetic-git-url> --provider git --scope user --agent claude --force
teamai status

Observed results:

• workspace status and pull exited 0 without Invalid local config;
• user-scope init exited 0 and emitted one warning for invalid recallEnabled: "yes";
• the invalid value was dropped;
• coAuthorEnabled: false, ai.tool: codex, and workflows.tools.default: codex were retained;
• the promoted config contained the user install record and remained readable;
• final status exited 0 without Invalid local config.

The related E2E fixture is skipped when Windows cannot create symlinks; its functional assertion runs where symlinks are available.

Disclosure

Slai Agent Runtime assisted with investigation, implementation, and test execution. The account owner reviewed and authorized this contribution and remains responsible for the submission.

Refs #1459

[github-actions]

• [P1 blocking] Missing real-CLI verification record — PR description, Validation. This PR changes runtime configuration behavior, but the description records only build, typecheck, lint, and unit tests; the E2E fixture is deferred to Linux CI. The first ## Code Review Rules item explicitly requires a completed end-to-end / real-CLI verification record. Run one representative settings-only user-init scenario after building and record the command and passing result.

No other confident findings in the specified diff.

[github-actions]

• [P1 blocking] Missing real-CLI verification record — PR description, Validation. The earlier finding remains unresolved. This runtime change requires a completed end-to-end / real-CLI record under the first ## Code Review Rules item; deferring the fixture to Linux CI is insufficient. Record one passing settings-only init --scope user run after building.

• [P3 nit] Run cleanup when skipping the symlink test — src/__tests__/config-not-initialized.test.ts:281. When symlink creation raises EPERM, the new return bypasses the existing finally, leaving HOME/USERPROFILE stubbed and the temporary home directory undeleted. Move symlink creation and its catch inside the cleanup-protected block.

No other confident findings in the specified diff. No PR code was run, built, or installed.

[huong8373tt-beep]

Addressed the requested real-CLI verification on the built dist/index.js.

I created an isolated temporary HOME and local bare Git remote, pre-seeded ~/.teamai/config.yaml with a settings-only file (no repo) containing:

ai:
  tool: codex
recallEnabled: true
workflows:
  tools:
    default: codex

Then ran:

teamai init <synthetic-git-url> --provider git --scope user --agent claude

The command exited 0 and reported teamai initialized successfully!. The resulting config contained the new user install record (repo, scope: user) and retained all pre-existing settings above. The explicit checks for repo_key, scope_user, ai_preserved, recall_preserved, workflow_tool_preserved, and no_config_error all returned true.

I also pushed ae3d8bb, which makes the existing symlink-only fixture return cleanly when Windows does not grant symlink creation permission; its functional assertion still runs wherever symlinks are available. The focused config suite now passes locally: 22 passed.

[github-actions]

• [P1 blocking] Missing real-CLI verification record — PR description, Validation. The earlier finding remains unresolved. This runtime change requires a completed end-to-end / real-CLI record under the first ## Code Review Rules item. Deferring the fixture to Linux CI is insufficient; record one passing settings-only init --scope user run after building.
• Resolved: The symlink-test cleanup finding is fixed in src/__tests__/config-not-initialized.test.ts:278; the EPERM return now runs the enclosing finally.

No other confident findings in the specified diff. No PR code was run, built, or installed.

[SaulMoro]

This was generated by AI during triage.

@huong8373tt-beep Thanks, this is step 1 of #1459, and the main part works. We ran it with the built CLI against main, with a settings-only ~/.teamai/config.yaml (no repo) beside a workspace-only or self-only install:

• No more config errors. status, doctor, pull and workflow cancel print no config error; on main they print 7, 29, 160 and 1 lines.
• Valid settings carry over. init --scope user over that file keeps every valid setting, asks nothing and sets nothing aside.
• Broken installs still fail. A file that fails validation, or has repo but no username, still reports could not be read.
• Mixed releases. 0.28.0 reads the promoted install.

Three changes before this is approved:

• [P1 blocking] init --scope user can write an install it can't read. With a machine file holding only recallEnabled: "yes" (a bare yes parses the same way), init --scope user exits 0. The next status then exits 1 with Invalid local config: recallEnabled: Expected boolean, received string, and the config could not be read. saveLocalConfig copies the file's keys into the new config without validating them; on main, init refused and left the file alone. #1459 gives the rule for a bad machine value: skip it, so the next level decides, and warn once. Please keep only the keys that validate, warn about each one dropped, and add a test with a wrong-typed value.
• [P1 blocking] Real-CLI record. This changes runtime behaviour, so the body needs one real-CLI run after npm run build (AGENTS.md, "Testing before a PR"). For example: a workspace install with a settings-only machine file, status and pull with no config error, then init --scope user keeping the settings.
• [P2 non-blocking, needed for approval] Closes #1459 should be Refs #1459. This is step 1 of the four in #1459's suggested order. Merging as is would close the issue with steps 2–4 still open.

Non-blocking:

• [P2 non-blocking] Every key of the file is carried into the new install, not only member settings, so hand-written install keys such as projectRoot or excludedSkills pass through. Carrying only the member-setting keys would match #1459.
• [P2 non-blocking] The kept settings are applied after init's first hook setup, so a machine codexTrustEnabled: false would not hold during the first init --scope user. We found this by reading the code; we didn't run it.
• [P3 nit] The two new commits make an existing Windows symlink test return on EPERM, so it passes without checking anything. Skipping it (ctx.skip()) would show it as skipped. It is also unrelated to this change, so it could go in its own PR.
• [P3 nit] One scope === 'user' guard is redundant, three comments still describe the old ai-only rule, and saveLocalConfig repeats writeLocalConfig's body.

Overlap: #1465 conflicts by one line in serializeLocalConfig; keeping both changes resolves it.

CI review: we agree that the real-CLI record is missing. It missed the wrong-typed value above.

Verified at 00012c4; edbc860 only changes that test: tsc and lint pass. config-not-initialized.test.ts passes (22 tests), also under a symlinked TMPDIR, and its two new tests fail on main. Nine related e2e files pass. Not run here: the codexTrustEnabled case, init --http, and Windows.

[huong8373tt-beep]

Thanks for the detailed review. Addressed the blocking invalid-value case in commit 3ca7b957.

init --scope user now validates each carried member setting independently: valid sibling settings are preserved, invalid values are skipped with a warning, and the promoted config remains readable. Added a regression test for recallEnabled: "yes" alongside a valid coAuthorEnabled value.

Validation after the fix:

• npm test -- src/__tests__/config-not-initialized.test.ts — 23 passed
• npm run build — passed
• npm run typecheck — passed
• npm run lint -- src/config.ts src/__tests__/config-not-initialized.test.ts — passed
• git diff --check — passed

The real-CLI settings-only init --scope user verification was also run against the built dist/index.js and recorded above. The PR continues to reference #1459 rather than close it, since this is step 1 of the four-part sequence.

[github-actions]

• [P1 blocking] Missing real-CLI verification record — PR description, Validation. The earlier finding remains unresolved. This runtime change requires a completed end-to-end / real-CLI record under the first ## Code Review Rules item. Deferring the fixture to Linux CI is insufficient; record one passing settings-only init --scope user scenario after building.
• Resolved: The symlink-test cleanup finding is fixed in src/__tests__/config-not-initialized.test.ts:278; the EPERM return now runs the enclosing finally.

No other confident findings in the specified diff. No PR code was run, built, or installed.

[huong8373tt-beep]

Follow-up to the review:

• PR1 remains Refs #1459 (not Closes) because this is step 1 of the four-part sequence.
• The invalid-value case is fixed in 3ca7b957; invalid machine settings are skipped individually while valid siblings are preserved.
• The requested real-CLI verification was run against the built dist/index.js with an isolated HOME and local bare Git remote; the passing scenario and retained settings are recorded above.
• The symlink-fixture cleanup was fixed in edbc860.

PR2 and PR3 are now rebased on this PR’s updated head and have also been force-pushed. The dependent sequence is therefore:

#1468 (3ca7b957) → #1473 (edad6259) → #1474 (f1aed7a5)

For PR2, validation includes the focused overlay suite (4 passed), build, typecheck, lint, and diff check. PR3 additionally validates doctor --json provenance against a built CLI: validated machine values, scope-over-machine precedence, merged model aliases, workflow defaults, and the local teamai.yaml provenance path.

[huong8373tt-beep]

Follow-up to the review:

• PR1 remains Refs #1459 (not Closes) because this is step 1 of the four-part sequence.
• The invalid-value case is fixed in 3ca7b957; invalid machine settings are skipped individually while valid siblings are preserved.
• The requested real-CLI verification was run against the built dist/index.js with an isolated HOME and local bare Git remote; the passing scenario and retained settings are recorded above.
• The symlink-fixture cleanup was fixed in edbc860.

PR2 and PR3 are now rebased on this PR’s updated head and have also been force-pushed. The dependent sequence is therefore:

#1468 (3ca7b957) → #1473 (edad6259) → #1474 (f1aed7a5)

For PR2, validation includes the focused overlay suite (4 passed), build, typecheck, lint, and diff check. PR3 additionally validates doctor --json provenance against a built CLI: validated machine values, scope-over-machine precedence, merged model aliases, workflow defaults, and the local teamai.yaml provenance path.

[SaulMoro]

This was generated by AI during triage.

Re-review at 3ca7b95. The wrong-typed value fix works. One blocking item is left.

Blocking

• [P1 blocking] The real-CLI record is still not in the PR description. The body's Validation section still says the e2e fixture "will run in Linux CI". The record is only in the 16:29Z comment, which is prose with no output and predates 3ca7b95. Please paste the built-CLI run into the body, with the commands and their output, including the recallEnabled: "yes" case. We ran it on our side and it behaves as described, so this is only the record.

Resolved since edbc860

Finding	Evidence
init --scope user with recallEnabled: "yes" wrote an install it could not read	Built CLI: init exits 0 and warns Ignoring recallEnabled in …/config.yaml, the key is dropped, valid siblings (coAuthorEnabled, ai.tool, workflows.tools.default) stay, the next status exits 0. The new unit test fails against edbc860's src/config.ts.
Every key was carried, not only member settings	Only the 11 member-setting keys from #1459 are carried now.
Closes vs Refs #1459	Refs kept.

Non-blocking, fits this PR

• [P2 non-blocking] Promotion now drops non-member keys without a word. A machine file with toolRoots, excludedSkills or gitExcludeEnabled loses them on init --scope user, with no warning. On main init refused and kept the file, and at edbc860 these keys were kept. toolRoots matters most, because once the file is promoted it is the user scope and #1459 gives toolRoots the rule "this scope, else this machine". Suggest one warning per dropped key, in the same form as the bad-value warning, or keeping toolRoots.
• [P3 nit] One bad entry drops the whole workflows block. workflows: {tools: {default: codex}, max_parallel: "lots"} keeps nothing, so the valid tools.default is lost. Per-entry handling fits #1473 too.
• [P3 nit] The warning prints Zod's (root): for a top-level value. #1459 suggests recallEnabled in ~/.teamai/config.yaml is "yes", not true or false, so it is ignored. The saveLocalConfig docstring still says it retains all settings.
• [P2 non-blocking] Still open from the last review: a kept codexTrustEnabled: false is applied after init's first hook setup. #1459 puts that in the overlay (#1473), so it can be closed there.

Verification at 3ca7b95

• npm ci, npm run build, npx tsc --noEmit, npm run lint exit 0.
• config-not-initialized.test.ts: 23 passed, also with a symlinked TMPDIR.
• e2e init-ends-with-pull, init-agent-validation, user-uninstall-keeps-projects-1025, init-unattended: all pass.
• Built CLI, isolated HOME and a local bare remote: a settings-only file next to a workspace-only and a self-only install (status, pull, doctor with 0 config errors), init --scope user over it, broken installs still rejected, a bad value, mixed keys, a bad workflows entry.
• With #1475 (fix for #1456) merged on top, a workspace pull over a settings-only machine file prints no config errors. #1475 alone prints Invalid local config on that pull, so this PR closes the caveat #1475 mentions.
• CI at 3ca7b95 is green. The red E2E (fork-safe) entry belongs to the run cancelled when the PR was closed.

[SaulMoro]

This was generated by AI during triage.

Thanks, @huong8373tt-beep. This change landed with #1473 (292e2af), which carried these commits and the built-CLI record for the settings-only machine file, so I'm closing this PR. #1474 and #1477 can rebase onto main now.