fix(dry-run): git previews for remove, roles/projects and import --from-repo (#900) by SaulMoro · Pull Request #971 · Tencent/teamai-cli

SaulMoro · 2026-10-03T15:30:10Z

Summary

 remove / roles init|add|update|remove / projects add|update|remove --dry-run
- pull/reset member clone, create knowledge worktree, save placement records
+ dirty dedicated clone -> refuse with commit/stash guidance, exit 1
+ clean clone -> fetch configured upstream -> model ff-only result
+ upstream ahead of HEAD -> preview upstream; HEAD ahead -> retain HEAD
+ failed pull / divergence / no upstream -> fetch origin/current, model reset
+ removal refresh failure -> refuse; manifest failure -> warn, read local HEAD
+ self mode -> read origin/default in a disposable checkout
+ removal placement records -> reconcile in memory without saving

 roles init (clone mode)
+ preview refresh -> existing-manifest check -> overwrite confirmation/rendering
  real pull -> existing-manifest check -> questions -> write (no second pull)

 import --from-repo / --from-repo-list --dry-run
- clone/fetch/reset cache -> lock -> AI scan
+ valid incremental cache -> query cached origin/branch
+ deleted branch + retained non-pruning origin ref -> read retained commit
+ pruning / explicit deleted-branch refspec / missing origin ref -> full clone
+ full clone or missing cache/LAST_SYNC -> query remote HEAD
+ report actual --output destination and cache freshness -> return without writes

 merged #969 guard
+ B preview entries restored; remote source precedence retained

Evidence

Before round 5: local release tracking origin/main or upstream/main fetched nonexistent origin/release before its configured upstream, so removal refused and manifests previewed stale contents. Public regressions also exposed stale upstream ancestry selecting the wrong divergence outcome (3 failing tests).
After: previews fetch the configured remote/ref first and select the same tree as real pullRepo. Both upstream naming cases pass for roles, projects and removal. Non-origin upstream works with origin unavailable. A distinct origin/release target is used only after upstream divergence; a no-upstream clone retains its reset fallback.
Earlier regressions remain green: pushed/ahead/behind/diverged feature branches; local-only branch refusal; dirty dedicated-clone refusal and dirty self-mode business files; failed-fetch policies; stale init overwrite refusal and single real pull; cached master versus remote HEAD; deleted cache branches, pruning and single-branch fallbacks; missing cache/LAST_SYNC; custom output and no cache/lock writes.

Fresh built CLI after build, isolated HOME: clean local release tracks upstream/main; origin is unavailable and the upstream has new role/rule contents.

$ node dist/index.js roles add remote-role --namespaces remote --dry-run
Role "remote-role" already exists. Use `teamai roles update remote-role` to modify it.
$ node dist/index.js remove rules fresh --dry-run
Will remove 1 rules: fresh
Dry run — no changes made
both exit 0; identical HEAD/branch/status/worktree list,
config/state/manifests/resources SHA-256 hashes before/after

$ node dist/index.js roles add remote-role --namespaces remote
Pulling latest changes... Up to date
Role "remote-role" already exists.
exit 0; real pull fast-forwards release to upstream/main, matching preview

Source fetch metadata and normal diagnostic logs are explicitly excluded from application-state snapshots. Import previews additionally verify all cache hashes including .git/LAST_SYNC.

Test Plan

npm run build
npx tsc --noEmit
npm run lint
npx vitest run: 388 files passed, 1 skipped; 8,021 tests passed, 20 skipped.
Public command / built CLI red-green regressions, including real pull comparisons.
Relevant E2E: 115 tests passed across 11 files: dry-run-git-previews-900, import-repo-dry-run-900, dry-run-refusal, remove-force, remove-mcp-namespaces, projects-manifest-crud, roles-tags-pull, role-scoped-agents, self-mode-worktrees-808, namespaced-entries and recall-feedback-dry-run.
Whole accumulated diff and shared-state readers/writers reviewed; git diff --check passed.

Hosted full E2E at prior head 0dd1ab51 failed five C preview cases because stats/recall guard entries are still refused. This is the integration work owned by open #977, not included or weakened here. See the hosted E2E job. Local B checks pass; hosted full-suite success depends on that integration.

Related Issues

Refs #900 (C1, C3, C5)

Merge Danger

Door: two-way

No migration. Reverting restores previous behavior.

Blast Radius: narrow

Git admin/removal previews model the real configured-upstream pull and origin/current reset fallback without moving the member checkout. Dirty clones refuse; self-mode default-branch behavior remains. Real import clone/fetch/reset semantics are unchanged.

Notes for Reviewers

Based on a8957a89, preserving merged fix(dry-run): refuse --dry-run on commands with no preview (#900) #969/fix(dry-run): stats and recall <query> dry runs write no local state (#900) #970/fix(rules): use native formats and scope-specific delivery #957/fix(code-knowledge): veto a Swift bare call only inside a type body #966 and fix(import): preview current organization repositories safely #960/fix(recall): honor --dry-run for feedback (#900) #965/fix(votes): read votes without migrating in viz and hasPendingVoteDeltas (#972) #973 behavior. B guard integration is implemented; no later restoration is needed for B.
C owns stats/recall guard/doc rows, config and session credits in fix(dry-run): restore stats and recall previews with inferred credits #977. Those rows and implementations are untouched here.
Clone init checks the refreshed temporary manifest before prompts; self-mode init remains unchanged.
Accepted C1 limitation remains: local scans read disk placement records while publishedNameFor receives reconciled in-memory records. No unrelated handler redesign.
Bilingual guide and B admin skill guidance updated. No help/flag or generated-reference changes. Extra providers/agents remain CI coverage.

github-actions · 2026-10-04T05:24:00Z

[P2 non-blocking] Do not silently use stale data when remove --dry-run cannot fetch. src/utils/reports-branch.ts:253 catches every fetch failure and continues from the last origin/<branch>. For example, if origin is unreachable and the stale ref still contains a deleted rule, the dry run reports that it would remove the rule and exits successfully, while the real non-self remove rejects the failed refresh and removes nothing. Propagate the failure for remove, or make stale fallback caller-configurable.
[P2 non-blocking] Avoid pulling again after roles init checks for an existing manifest. src/manifest-edit.ts:35 now performs a second pull when rolesInit calls it at src/roles-cmd.ts:128. If another administrator adds manifest/roles.yaml while the user is answering the interactive questions, this pull retrieves it and the callback immediately overwrites it without the overwrite confirmation at src/roles-cmd.ts:38. Skip the second pull for initialization or repeat the existence check afterward.

The PR description includes sufficient real-CLI end-to-end verification for this runtime change.

SaulMoro · 2026-10-04T09:35:59Z

Addressed both findings in 9703fcb after rebasing onto e465eec. Clone-mode remove --dry-run now refuses a failed fetch with the real removal's message and exit 1; stale fallback is explicit for roles/projects and self-mode removal. roles init skips the second pull, preserving the checkout checked before the prompts. Both regressions failed before the fix and pass now. Build, tsc, lint, 7,644 unit tests and 79 affected e2e tests pass. The built unreachable-origin CLI run leaves application state unchanged, with normal diagnostic logging excluded. The PR body includes the evidence and later-merging #969 guard integration. Reviews have not been resolved or dismissed.

github-actions · 2026-10-04T09:43:11Z

Findings

[P2 non-blocking] Preserve local manifest edits that a real pull would retain. withDefaultBranchPreview always checks out origin/<default>, but pullRepo --ff-only can succeed while leaving uncommitted manifest changes intact. For example, if manifest/roles.yaml contains an uncommitted role and origin is unchanged, roles add <same-role> --dry-run reports it can add the role, while the real command reports it already exists. Either refuse dirty clones or construct the preview from the actual post-pull working tree. src/utils/reports-branch.ts:267
[P2 non-blocking] Compare incremental caches against the branch they will actually refresh. The preview queries remote HEAD, while the real incremental path resets origin/${currentBranch}. If a repository changes its default branch from master to main, the preview reports the new main SHA, but the real run continues updating master and may report the cache unchanged. Query the cached checkout’s branch for incremental previews, or update the real path to follow remote HEAD. src/import-repo.ts:219

Review Status

Both findings from the earlier review are resolved.
The PR description includes sufficient representative real-CLI end-to-end verification.

SaulMoro · 2026-10-04T09:59:22Z

Addressed this review in 59f40d8. Dedicated-clone previews now refuse uncommitted changes with commit/stash guidance and exit 1, preserving dirty files and HEAD. Dirty self-mode business files remain supported, and prior refresh policies are unchanged. Incremental import previews query the cached branch at its configured origin using fetch authentication; full-clone and missing-cache/LAST_SYNC previews follow remote HEAD. Both reported cases failed before the fix and pass now, including a real incremental command confirming unchanged master after remote HEAD moved to main. Build, tsc, lint, 7,644 unit tests and 87 relevant e2e tests pass. Built CLI records verify unchanged clone/cache hashes, including .git, with normal diagnostic logging excluded. Body updated before normal push; reviews left untouched.

github-actions · 2026-10-04T10:05:20Z

Findings

[P2 non-blocking] Preview the same branch that clone-mode commands actually edit. withDefaultBranchPreview always checks out origin/<default>, while the real pullRepo operates on the currently checked-out branch. After a push succeeds but PR creation fails, the clean team clone remains on its pushed feature branch; a subsequent roles, projects, or remove --dry-run previews default-branch contents while the real command reads and modifies the feature branch. src/utils/reports-branch.ts:272
[P2 non-blocking] Refresh roles init --dry-run through a temporary checkout before checking for an existing manifest. A stale clone without manifest/roles.yaml will currently proceed and print a new manifest even when the file already exists upstream, whereas the real command pulls first and requests overwrite confirmation. src/roles-cmd.ts:34
[P3 nit] Honor --output in the import preview destination. The message always reports teamwiki/evidence/code/<slug>, although a real run with --output derives a different teamwikiRoot, so the preview reports the wrong target location. src/import-repo.ts:237

Review Status

Both findings from the first review and both findings from the second review are resolved.
The PR description includes sufficient representative real-CLI end-to-end verification.

jeff-r2026 · 2026-10-04T12:45:31Z

This branch has merge conflicts with main. Please rebase onto the latest main and resolve the conflicts so review can continue.

…nt#900) roles add/remove/update and projects add/update/remove pulled the team repo before their dry-run guard, and a pull that cannot fast-forward resets the clone to origin, discarding unpushed commits. In single-repo mode they also created the knowledge worktree. runManifestEdit now takes { dryRun }: a preview fetches and reads origin/<default> from a --shared clone in the OS temp dir (withDefaultBranchPreview), so the member's repo only sees a fetch. roles init skips its pull on a dry run. Refs Tencent#900 (C5)

…#900) remove pulled the team repo and saved the reconciled placement records before its dry-run guard; in single-repo mode it ran inside the knowledge worktree. A dry run now resolves names in a temporary checkout of origin/<default>, as the real run does after its pull, and hands the reconciled records to publishedNameFor in memory instead of saving them for it to read back. Refs Tencent#900 (C1)

A dry run of import --from-repo and --from-repo-list deleted and re-cloned the repo cache (or fetched and reset it), took the import lock and ran the LLM scan. It now reads the remote head with git ls-remote, using the clone's auth (resolveCloneAuth, extracted from shallowClone), compares it with LAST_SYNC and prints "Would import <owner>/<repo> at <sha>" with the cache state, then stops. Refs Tencent#900 (C3)

…nt#900)

…encent#900)

…ncent#900)

SaulMoro · 2026-10-04T14:28:31Z

Addressed the three findings from #971 (comment) in 7ffe273. Clean clone previews now match the checked-out branch's real ff-only pull/reset outcome (pushed, ahead, behind and diverged); self mode still reads origin/default and dirty dedicated clones still refuse. Clone roles init --dry-run checks the refreshed temporary manifest before overwrite confirmation, and import preview honors the actual --output destination without writes. The six public regressions failed before the fixes and pass now. B's preview guard entries and remote-source precedence are restored after #969. Build, tsc, lint, 8,021 unit tests and 106 relevant E2E tests passed; a built real CLI run matched an ahead feature-branch pull with unchanged HEAD, branch, status, worktree list and application hashes. Extra providers/agents remain CI coverage. PR body includes the evidence; reviews were not dismissed or resolved.

SaulMoro · 2026-10-04T14:28:33Z

Rebased onto current main a8957a8 and pushed 7ffe273, preserving the merged #969/#970/#957/#966 changes. The CHANGELOG conflicts are resolved. Build, typecheck, lint, full unit tests and relevant E2E checks passed. This addresses the rebase request at #971 (comment).

github-actions · 2026-10-04T14:32:19Z

Findings

[P2 non-blocking] Fetch the active branch explicitly for removal previews. On a clean local-only branch with no origin/<branch>, git fetch origin succeeds and the preview uses local HEAD, but the real removal’s git fetch origin <branch> fails and refuses to remove anything. The dry run therefore reports an operation that cannot run. src/utils/reports-branch.ts:283
[P2 non-blocking] Match incremental behavior when the cached branch was deleted remotely. remoteHeadSha throws for the missing branch and this catch previews a full clone of the new default branch, while the real non-pruning shallowFetch can retain and reset to the stale origin/<branch>. For example, after deleting master and moving HEAD to main, preview and execution inspect different commits. src/import-repo.ts:227

Review Status

The previously reported findings are resolved in their stated scenarios.
The PR description includes sufficient representative real-CLI end-to-end verification.

SaulMoro · 2026-10-04T14:45:31Z

Addressed both findings from #971 (comment) in 0dd1ab5. Removal preview now explicitly fetches the active branch, so a local-only branch produces the same refresh refusal and exit 1 as real removal, with HEAD/files/state untouched. Incremental import preview uses a deleted branch's retained origin ref when the real non-pruning fetch/reset retains it; pruning, explicit deleted-branch refspec and missing origin-ref cases keep the full-clone fallback. Real import semantics are unchanged. Both public regressions failed before the fixes and pass now, including comparisons with real CLI execution and public shallowFetch fallback checks. Build, tsc, lint, 8,021 unit tests and 111 relevant E2E tests passed. The PR body includes fresh built-CLI evidence and cache/application-state hash checks. Extra providers/agents remain CI coverage; no reviews were dismissed or resolved.

github-actions · 2026-10-04T14:49:58Z

Findings

[P2 non-blocking] Follow the configured upstream before treating the active branch as unrefreshable. git fetch origin <local-branch> fails when a local branch tracks a differently named remote branch—for example, local release tracking origin/main. The real pullRepo() successfully fast-forwards from origin/main, but remove --dry-run refuses and manifest dry runs use stale local contents. Fetch/resolve @{upstream} first, then use origin/<branch> only when modeling the hard-reset fallback. src/utils/reports-branch.ts:283

Review Status

All previously reported findings are resolved in their stated scenarios.
The PR description includes sufficient representative real-CLI end-to-end verification.

…nt#900)

SaulMoro · 2026-10-04T15:33:22Z

Addressed #971 (comment) in 20d68c4. Clone previews now fetch the configured remote/merge ref first and model its ff-only result; origin/current-local-branch is used only for the real pull/reset fallback. Public red/green tests cover release tracking origin/main, a non-origin upstream with origin unavailable, a distinct reset target after divergence, and no-upstream fallback. Roles, projects and removal read the same post-pull tree; earlier dirty/self-mode/local-only/import regressions remain green. Build, tsc, lint, 8,021 unit tests and 115 relevant E2E tests passed. Fresh built CLI evidence shows preview parity with a real upstream/main pull and unchanged dry-run HEAD/files/state. Prior hosted E2E's five C guard failures remain the integration dependency owned by #977; C rows and tests are untouched. PR body includes the evidence and dependency status. No reviews were dismissed or resolved.

github-actions · 2026-10-04T15:40:06Z

Findings

[P3 nit] Avoid exposing the disposable checkout path in manifest diagnostics. runManifestEdit passes the temporary repository path into callbacks, so roles init --dry-run reports an existing manifest under /tmp/teamai-preview-*, and other role previews can report a missing manifest there. That directory is immediately deleted, making the guidance misleading; use the configured repository path for user-facing messages. src/manifest-edit.ts:44

Review Status

All previously reported findings are resolved in their stated scenarios.
The PR description includes sufficient representative real-CLI end-to-end verification.

This was referenced Oct 3, 2026

[bug] --dry-run still writes: recall maintenance/promote run for real, writes past the loader, and flags that are ignored #900

Open

fix(import): preview current organization repositories safely #960

Merged

jeff-r2026 assigned SaulMoro and jeff-r2026 and unassigned SaulMoro Oct 4, 2026

SaulMoro mentioned this pull request Oct 4, 2026

fix(dry-run): refuse --dry-run on commands with no preview (#900) #969

Merged

7 of 11 tasks

SaulMoro force-pushed the fix/dry-run-git-previews branch from 23f026e to 9703fcb Compare October 4, 2026 09:35

SaulMoro added 5 commits October 4, 2026 16:00

fix(dry-run): preserve refresh refusal and roles init ordering (Tence…

bd6adc7

…nt#900)

fix(dry-run): refuse dirty clones and preview cached import branch (T…

bd3a418

…encent#900)

SaulMoro mentioned this pull request Oct 4, 2026

fix(dry-run): restore stats and recall previews with inferred credits #977

Open

1 task done

fix(dry-run): match post-pull branches and restore preview guards (Te…

7ffe273

…ncent#900)

SaulMoro force-pushed the fix/dry-run-git-previews branch from 59f40d8 to 7ffe273 Compare October 4, 2026 14:27

fix(dry-run): match missing remote branch refresh outcomes (Tencent#900)

0dd1ab5

fix(dry-run): follow configured upstream before reset fallback (Tence…

20d68c4

…nt#900)

Navigation Menu

fix(dry-run): git previews for remove, roles/projects and import --from-repo (#900) - #971

fix(dry-run): git previews for remove, roles/projects and import --from-repo (#900)#971
SaulMoro wants to merge 8 commits into
Tencent:mainfrom
SaulMoro:fix/dry-run-git-previews

SaulMoro commented Oct 3, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

jeff-r2026 commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

SaulMoro commented Oct 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Evidence

Test Plan

Related Issues

Merge Danger

Notes for Reviewers

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

jeff-r2026 commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

SaulMoro commented Oct 4, 2026

Uh oh!

github-actions Bot commented Oct 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

SaulMoro commented Oct 3, 2026 •

edited

Loading