fix(cli): scaffold output survives a clean clone and a non-interactive run

[rickylabs]

Summary

Grouped beta.12 fix: everything netscript init emits should survive a clean clone and a
non-interactive run.

Four issues, one verification path — scaffold a project without a TTY, clone it clean, start it:

• fix(scaffold): generated .gitignore excludes appsettings.json, which a clean clone needs #966 — the generated .gitignore excludes appsettings.json, which is source configuration
  (aspire/apphost.mts reads it, netscript generate derives from it, nothing regenerates it).
  A clean clone cannot start.
• fix(aspire): scaffold writes a top-level 'Parameters' block the appsettings schema drops #975 — generateAppsettings() writes a top-level Parameters block that AppSettingsSchema
  does not model, so parseAppSettings() silently strips it and config set demands --force.
• fix(cli): init nests a project directory when the cwd is already the target #967 — init <name> from a directory already meant to be the workspace nests <name>/
  inside it instead of initialising in place.
• fix(cli): non-interactive invocations still hit selection prompts #968 — non-interactive invocations (CI, agents, scripts) hit selection prompts and fail
  instead of defaulting or naming the flag that would have answered them.

Status

Draft. Implementation is delegated to a Codex slice on this branch; the run dir is
.llm/runs/fix-scaffold-hygiene--966/. This PR body will be filled in with root cause, the
regression guard and its fails-before evidence, and scoped gate output before it leaves draft.

Closes #966
Closes #975
Closes #967
Closes #968

[rickylabs]

IMPL-EVAL verdict — BLOCKED, staying in draft

The implementation slice wrote a coherent fix but never committed or pushed it. origin/fix/scaffold-hygiene is at 361a5e767, identical to local HEAD; both commits on this branch are run artefacts (context-pack.md, implement.md). git diff origin/main...HEAD touches zero source files. All 13 modified + 4 untracked files live only in the worktree /home/codex/repos/b12-scaffoldhygiene.

Cause: the pre-commit asset-barrel check exited 1 over the intended packages/cli/src/kernel/assets/embedded.generated.ts diff, and the agent turn ended without retrying. The work is on disk and looks correct — it is one regenerate-and-commit away — but nothing is on the remote, so it is one git checkout . from being lost.

The four evaluator checks

1. Gate evidence covers the changed files — CANNOT BE ASSESSED. There is no committed diff to gate. Every PASS in the implementation report is Codex's self-report from worklog.md, not an observed run against a pushed tree. The one claim I did re-run independently, the focused regression suite, gives 24 passed / 36 steps, not the reported 25 / 37 — a small but real discrepancy in the self-report. asset_barrel_check is NOT RUN (it is the blocker) and scaffold.runtime E2E is NOT RUN.

2. Root cause vs symptom — the grouping hypothesis DID NOT HOLD. This group existed on the belief that four issues shared one cause. They do not. The slice landed four independent patches in four unrelated places:

• gitignore.template — drop the appsettings.json exclusion (fix(scaffold): generated .gitignore excludes appsettings.json, which a clean clone needs #966)
• AppSettingsZod / AppSettings — model a Parameters record (fix(aspire): scaffold writes a top-level 'Parameters' block the appsettings schema drops #975)
• validate-init.ts — basename(cwd) === name ? cwd : resolve(cwd, name) (fix(cli): init nests a project directory when the cwd is already the target #967)
• init-command.ts ×2 + init-input.ts — add a --non-interactive flag that maps to the existing ci path (fix(cli): non-interactive invocations still hit selection prompts #968)

Codex was explicit that it rejected the single-cause hypothesis, naming two contract clusters instead (generated-artifact contract, invocation contract). I agree with the finding, and it is honest work — but it should be recorded plainly: this is four symptom-local patches, not one root-cause fix. #966 and #975 are genuinely adjacent (both are "the scaffold's contract disagrees with what it emits"); #967 and #968 share nothing with them or with each other beyond touching init.

3. The issue was wrong — CONFIRMED, and the correction is properly posted. #968's filed cause was stale: main has bypassed init prompts on non-TTY stdin since beta.5. The correction is a real comment on the issue itself (issues/968#issuecomment-5144895872), not buried in a PR body. This is the one check that fully passes. Consequence worth flagging: what remains of #968 is a discoverability change, not a defect fix — it adds a --non-interactive spelling for behaviour that already worked.

4. Regression guard fails-before — PARTIAL, and #968 has no real guard. Guards exist for #966 (plan-init_test.ts), #967 and #968 (init-command_test.ts), #975 (config_test.ts, schema_test.ts). Codex's own pre-implementation run reports 19 passed / 5 failed, which is credible fails-before evidence for #966, #967 and #975. But the #968 guard PASSED before the fix — by definition it is not a regression guard for the filed defect; it is a characterisation test for behaviour that was already correct. I did not run the revert-and-confirm-fails loop myself, because there is no committed hunk to revert.

What Eric must decide

The recovery is cheap and the diff on disk is good. Recommended: resume Codex thread 019fb8e4-f036-7290-bff2-0d140ec053e2 in the worktree with a single instruction — regenerate the asset barrel, commit, push, verify with git ls-remote. Do not relaunch the implementation; a fresh launch assumes a clean tree and would discard the work.

Two things to weigh at merge time:

• Whether fix(cli): non-interactive invocations still hit selection prompts #968 should stay in this PR's Closes list at all, given its stated cause was wrong and its remaining content is a flag alias. It may be more honest to close fix(cli): non-interactive invocations still hit selection prompts #968 with the correction comment and drop it from this PR.
• packages/aspire JSR audit fails on 4 pre-existing missing @module tags (application, adapters, testing, public). Pre-existing baseline, untouched by this slice, tracked as drift D2 — not a blocker for this PR but it is not clean.

Branch is also 1 commit behind origin/main (37047e283) and will need a rebase before merge.

PR stays draft and keeps status:in-progress until the implementation is actually on the remote.

[rickylabs]

Supervisor recovery + verification

The implementation was complete but stranded in the worktree — 13 modified and 4 untracked
files, never committed. The cause is in the slice's own worklog: check:assets-barrel runs
gen:assets-barrel && git diff --exit-code, so the intended generated diff in
embedded.generated.ts (from #966's .gitignore template change) made the gate exit 1, and the
turn ended there. Regenerated the barrel, committed, pushed — the gate passes now that the diff is
committed, which confirms the diagnosis rather than working around it.

Also merged origin/main (the branch was behind 37047e283). One conflict, in
validate-init.ts: this branch added basename for #967's cwd-is-already-the-target check while
#978 added USER_PORT_RANGE for widened host-port validation. Both are used; resolution is the
union.

One pre-existing test changed deliberately

resolve-appsettings-path_test.ts asserted Parameters.postgres-password resolves as unknown —
true only because of the #975 defect. That test carried two assertions and only the status was
stale; the one that mattered (a top-level key is not silently re-homed under NetScript) is
untouched. The undeclared-key case is preserved as its own test against a key the schema genuinely
does not declare, so widening the schema did not quietly retire the guard that the widening was
scoped.

Gates — scoped, because root lint/fmt:check exclude packages/cli

Gate	Result
Tests (packages/cli + packages/aspire)	497 passed / 518 steps, 0 failed
Lint (scoped to both roots)	0 occurrences
Format (scoped to both roots)	0 findings
check:assets-barrel	PASS (clean once the generated diff is committed)
CI on 382d2795	all green, incl. scaffold-runtime and close-gate

close-gate passes because #975's two acceptance criteria were verified against the diff and ticked
with evidence on the issue — not ticked to make the gate go quiet.

Two things for Eric, both honest limitations

1. The grouping hypothesis did not hold here. These are four symptom-local patches in four
   unrelated places (gitignore template, AppSettingsZod, target resolution in validate-init.ts,
   a --non-interactive alias) — not one shared-cause fix. Only fix(scaffold): generated .gitignore excludes appsettings.json, which a clean clone needs #966 and fix(aspire): scaffold writes a top-level 'Parameters' block the appsettings schema drops #975 are genuinely
   adjacent. Worth recording, because the same grouping did pay off on the plugin cluster (fix(plugin): generated plugin glue reaches its runtime #988),
   so this is evidence about which groupings are real rather than a failure of the method.
2. fix(cli): non-interactive invocations still hit selection prompts #968's regression guard is a characterisation test, not a guard — it passed against the
   broken code. Its filed cause was stale (main has bypassed non-TTY prompts since beta.5,
   corrected at fix(cli): non-interactive invocations still hit selection prompts #968#issuecomment-5144895872), so what remained was the missing --non-interactive
   spelling. Whether fix(cli): non-interactive invocations still hit selection prompts #968 should stay in this PR's Closes list is your call — closing it via
   the correction alone may be more honest than closing it via a flag alias.

Not merged.

[augmentcode]

🤖 Augment PR Summary

Summary: Hardens the netscript init scaffold output so a clean clone (including non-interactive/CI runs) can start reliably.

Changes:

• Stops ignoring appsettings.json in the scaffolded .gitignore, keeping required source config tracked (fix(scaffold): generated .gitignore excludes appsettings.json, which a clean clone needs #966).
• Extends Aspire AppSettings / Zod schema (and generated JSON Schema) to model top-level Parameters, so the scaffolded config is preserved and schema-aware tooling can set it without --force (fix(aspire): scaffold writes a top-level 'Parameters' block the appsettings schema drops #975).
• Adjusts init target-path resolution so netscript init <name> initializes in-place when the CWD basename already matches <name> (fix(cli): init nests a project directory when the cwd is already the target #967).
• Adds an explicit --non-interactive flag (mapping to existing CI behavior) for both public and maintainer init commands (fix(cli): non-interactive invocations still hit selection prompts #968).
• Adds focused regression tests for the above contracts (gitignore tracking, schema emission, config set path resolution, and non-terminal prompt bypass).

Technical Notes: Non-interactive resolution is centralized by mapping --non-interactive to ci before calling the interactive input resolver; schema changes flow through to generated JSON Schema tests.

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. 2 suggestions posted.

Comment augment review to trigger a new review at any time.

[augmentcode]

packages/cli/src/kernel/application/scaffold/plan-init_test.ts:191 — This assertion only checks for an exact appsettings.json line, so it would miss ignore patterns like /appsettings.json or **/appsettings.json and may not reliably guard #966 regressions.

Severity: low

_{🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.}

[augmentcode]

packages/aspire/tests/schema_test.ts:39 — JSON.stringify(schema).includes('Parameters') can pass even if Parameters isn’t actually a top-level property (e.g., the word appears in a description/$defs), so this guard may be weaker than intended.

Severity: low

_{🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.}