Files
kandev/docs/specs/platform/system-design/workspace-git-status.md
T
2026-09-13 21:57:54 +01:00

30 KiB

status, system, requirements, created, updated, owners
status system requirements created updated owners
draft platform
REQ-PLATFORM-WORKSPACE-GIT-STATUS-001
2026-07-19 2026-08-31
kandev

Workspace Git Status System Design

Purpose and boundaries

This design preserves the technical source detail for REQ-PLATFORM-WORKSPACE-GIT-STATUS-001 during migration.

Requirement mapping

Requirement Design section
REQ-PLATFORM-WORKSPACE-GIT-STATUS-001 Migrated source detail, Generated untracked dependency trees, Mixed index and working-tree changes, Comparison-target Git execution, API surface

Migrated source detail

Why

Users opening or focusing Changes and Review need a current workspace snapshot without a large generated or untracked tree monopolizing agentctl. Repeated requests for the same repository must not amplify expensive Git and filesystem work, and the initial session-hydration path must remain within its two-second live-status budget by falling back when necessary.

What

  • Cached reads return the latest workspace-tracker snapshot. When no cached snapshot exists, the tracker performs a live observation.
  • Fresh reads observe the live worktree and do not themselves replace the polling cache.
  • Overlapping live observations for the same repository share one underlying observation. Different repositories in a multi-repository task may still be observed in parallel.
  • Every non-cancelled caller receives the same completed snapshot or error from a shared observation. A caller whose own context is cancelled returns promptly without cancelling or otherwise poisoning the result for other callers.
  • Tracker shutdown or the bounded shared-observation deadline cancels the underlying work. Cancelled work does not publish or cache a partial snapshot.
  • After Git output is parsed, changed-file and synthetic untracked-diff enrichment performs work proportional to the number of eligible changed entries plus the bounded content processed.
  • Existing diff limits remain in force: 10 MiB maximum source file size, 256 KiB maximum per emitted diff representation, and a 2 MiB enrichment threshold per status snapshot. Every flattened or layer-specific representation participates in that threshold. Because the threshold is checked before enriching each representation, the final accepted representation may preserve the existing overshoot of up to the 256 KiB cap. Existing skip reasons remain unchanged.
  • Large changed sets retain every eligible path and its status metadata. Once the total diff budget is exhausted, files that are not enriched retain budget_exceeded as their diff skip reason.
  • Multi-repository responses retain repository identity and partial-success behavior.
  • Verification tooling preserves shared managed Go and lint caches for reuse while keeping invocation scratch and command output outside repository worktrees. The root-level .verify-cache and .tmp paths are ignored as safeguards against legacy or misconfigured verification runs.

Generated untracked dependency trees

Agentctl owns the complete workspace snapshot. It collects tracked and untracked state through separate Git queries so a generated dependency tree cannot force Git to emit every package file.

The tracked query uses git status --porcelain --untracked-files=no. It preserves index and working-tree changes for every tracked path, including a tracked path below node_modules. The untracked query uses git ls-files --others --exclude-standard --exclude=node_modules/ -z. Git applies the extra exclusion while it walks the worktree. The -z result preserves path names without quoting or line splitting. Agentctl adds each returned path to the existing untracked status projection before diff enrichment.

For each full observation, agentctl pins both queries to a temporary snapshot of the Git index. Git writes the real index atomically, so the snapshot keeps tracking membership stable across the pair without blocking concurrent user Git operations. The snapshot is removed after the observation.

The extra exclusion applies to directories named node_modules at any repository depth. It also covers pnpm's nested node_modules/.pnpm store. It does not cover .next, dist, build, or other generated names. Those paths continue to follow Git's repository, global, and command-level ignore rules.

The full status observer and the untracked monitor fingerprint use one shared argument definition. The monitor therefore does not stat dependency-tree files, and changes limited to an untracked dependency tree do not start a full refresh. Filtering parsed status output is not sufficient because Git and agentctl have already paid the enumeration cost at that point.

The status model, HTTP routes, WebSocket events, and frontend stores do not change. A completed polling refresh replaces any older cached snapshot through the existing tracker path. This contract follows ADR-2026-08-30-bound-untracked-dependency-enumeration.

Mixed index and working-tree changes

Git porcelain reports two independent status columns for each path: the index state (X) and the working-tree state (Y). A path such as MM README.md or AM new-file.txt is one file identity with two change facets. The workspace model MUST preserve both columns instead of selecting one as the file's only state.

GitStatusUpdate.Files remains keyed once by repository-relative path. FileInfo retains its flattened fields as a compatibility projection and adds optional staged_change and unstaged_change objects. Each object carries the facet's status, additions, deletions, old path, diff, and diff skip reason. Agentctl emits the facet objects for a mixed path; a path with only one layer continues to use the existing flattened representation. Consumers MUST accept both shapes.

For a mixed path, agentctl computes three intentional views:

  • the flattened compatibility view compares HEAD with the working tree and preserves the existing worktree-priority staged=false classification for older consumers;
  • staged_change compares HEAD with the index using git diff --cached; and
  • unstaged_change compares the index with the working tree using git diff without a base ref.

All serialized diff strings, including the compatibility projection, participate in the existing snapshot-wide enrichment budget. Each representation retains the existing output cap and skip reasons. Budget exhaustion may omit one representation without discarding either facet's status metadata.

The frontend keeps the raw file map unique by (repository_name, path) for task badges, Review aggregation, mutation routing, and change-file totals. Changes derives two temporary FileInfo views from a mixed entry, one per facet, and falls back to the flattened fields when facet objects are absent. Stage, unstage, and discard operations continue to send only the repository-qualified path, never the temporary view identity.

Diff navigation adds the change layer to its target identity. Desktop preview and pinned diff tabs use (repository_name, path, layer) so the staged and unstaged views of one path cannot alias each other. The mobile full-height diff drawer carries the same layer while retaining its existing sheet, touch-target, and scroll behavior. Review's all-files source remains a single combined uncommitted file; only a layer-qualified Changes selection projects a facet-specific diff.

Status equality, editor synchronization signatures, and Changes focus fingerprints include the optional facet fields. A snapshot that changes only one facet must therefore publish and render. This contract follows ADR-2026-08-27-mixed-git-change-facets.

Base-commit staleness and refresh

The commits panel (git log <base>..HEAD) and cumulative diff anchor to a base commit derived from the session's stored base_commit_sha and base_branch. That anchor becomes stale when the branch history moves relative to the true integration branch after the base was recorded — most commonly a stacked-PR parent that merges and disappears, a rebase onto the integration branch, or a base branch that was deleted upstream. When the anchor is stale, the panel enumerates commits that are not part of the branch's own contribution, inflating the count.

  • A stored base commit SHALL be treated as stale when it is a strict ancestor of the true merge-base between HEAD and the resolved integration candidate — the first branch that resolves through the priority list origin/main → origin/master → main → master. When merge-base(HEAD, <resolved integration candidate>) advances past the stored base_commit_sha, the stored base is stale. A stored base that equals or is a descendant of that merge-base is NOT stale.
  • Staleness correction SHALL be limited to comparison targets that are not a live non-default upstream branch. The resolved default integration target (main/master, or its origin/* form) remains subject to the normal stale-base check. When the target IS a live non-default upstream branch (for example a feature comparing against origin/develop that has merged newer origin/main commits), its own merge-base is authoritative and SHALL NOT be overridden merely because it is an ancestor of the integration merge-base — doing so would hide commits that are genuinely part of that comparison.
  • When the stored base is stale, commit enumeration and cumulative diff SHALL use the freshly computed merge-base against the resolved integration candidate instead of the stored SHA. The panel therefore reflects only the commits the branch actually introduces over the integration branch.
  • Base resolution SHALL prefer the upstream integration ref (origin/<name>) over a bare local ref of the same name. A local ref that no longer tracks any live upstream (for example a merged/deleted stacked-parent branch that lingers only as a local ref) SHALL NOT anchor the commit range when an upstream integration ref is available.
  • Staleness detection is a read-time correction: it changes which base the commits/diff are computed against. It does not by itself rewrite the persisted base_commit_sha; the persisted value continues to follow the existing capture and the "Compare against" base-branch reset paths.
  • When neither an upstream integration ref nor a usable merge-base can be resolved (unrelated histories, offline mirror with no origin/*), behavior falls back to the existing stored-base / branch-tip anchor and the panel is unchanged from today.

Repository-qualified comparison targets

A branch name alone is not a complete comparison identity. A task may be attached to a contributor fork while its pull request targets a branch with the same name in another repository. In that case origin/main is the fork's main, not the pull request base.

  • Each task-repository attachment MAY persist one versioned, credential-free comparison_target in its metadata. The target identifies the provider change, the validated head repository and branch, and the target repository and branch.
  • The attachment row is the scope boundary. A target applies only when the provider-reported head repository matches the attached repository and the normalized head branch matches that attachment's live checkout branch. Repository-only matches, historical pull requests, and sibling branches MUST NOT retarget the attachment.
  • When the target repository is the attached repository, the existing origin/<base_branch> behavior remains authoritative and no extra remote is required.
  • When the target repository differs, Kandev materializes a deterministic comparison-only remote without renaming or rewriting origin, fetches only the validated target branch, and compares against that exact remote-tracking ref.
  • The repository-qualified ref is authoritative for task-card Git totals, ahead/behind, commits, cumulative diff, and Review expansion. Once configured, those surfaces MUST NOT silently fall back to origin/<branch> or a bare branch with the same name.
  • A successful pull-request association or retarget updates the matching attachment, clears its sessions' stored base commit, refreshes the live agentctl comparison, and invalidates client commit/diff caches. The corrected result appears without a session restart. Launch and resume reconstruct the target from attachment metadata.
  • Explicitly choosing a branch in Compare against returns the attachment to repository-local comparison: the write updates base_branch and removes any provider-derived comparison_target atomically.
  • A detached pull-request association removes the target only when that target names the detached change. Closing or merging the associated change does not remove the target because its base remains the relevant integration line for the task history.
  • Kandev does not broaden task Git credentials to materialize a target. Public targets and targets already readable through the execution's effective Git credentials are supported. An unreadable private target fails visibly.
  • Target setup is fail-closed. If identity validation, remote collision checks, fetch, or merge-base resolution fails, file status remains usable, but comparison-derived totals, ahead/behind, commits, and cumulative diff are unavailable. The Changes panel names the intended <owner>/<repository>:<branch> target and shows an actionable comparison error; task-list summaries suppress numeric Git totals and expose the unavailable state instead of publishing a same-named origin fallback.

Comparison-target Git execution

The instance's effective Git environment is authoritative for Kandev-owned workspace Git commands. This environment contains executor credentials and managed credential-helper entries. The process manager gives a detached environment copy to each workspace tracker. This rule applies to root, repository, submodule, rescan, and lazy trackers.

Each tracker starts Git with its instance environment. Final execution preparation follows Managed Git execution, including deny-only askpass, SSH option preservation, finite execution budgets, and bounded helper cleanup. The shared boundary applies controls after caller environment assembly. Tracker-specific index selection and optional-lock suppression remain owned by the tracker.

The process manager sets an explicit target to pending before it starts network work. It then materializes the target with the manager's lifetime context. Initial materialization, live updates, rescan, and lazy tracker creation do not wait for the fetch. A successful fetch publishes the exact comparison ref and starts a detached status refresh. An error publishes a bounded unavailable state. Manager shutdown cancels unfinished materialization.

Kandev selects the transport before it starts a comparison-target command. Canonical comparison targets remain HTTPS. An HTTPS authentication or transport error does not start an SSH retry. This rule prevents a command from changing identity, host trust, or credential scope after the first error. Kandev does not rewrite origin, the checkout upstream, or push routing.

An explicit fresh status request may re-evaluate an unavailable comparison target once. A moving target refresh may force-update only its deterministic internal comparison ref, so a stale cache does not block recovery; user remotes and push routing are unchanged.

This contract follows ADR-2026-08-31-deterministic-noninteractive-git-transport.

API surface

Existing Git-status routes remain in place. Their result and stream payloads add optional comparison state so old clients continue to decode them:

  • GET /api/v1/git/status?repo=<subpath>&fresh=<bool> returns the existing GitStatusResult shape.
  • GET /api/v1/git/status/multi?fresh=<bool> returns the existing MultiRepoGitStatusResult shape containing PerRepoGitStatus entries.
  • The fresh query parameter continues to select a live observation rather than a cached tracker snapshot.
  • A repository status MAY include comparison_target with its credential-free display identity and comparison_error_code when the explicit target is unavailable.
  • A mixed-path FileInfo MAY include staged_change and unstaged_change. Each facet uses the same bounded status, line-total, rename, diff, and skip-reason vocabulary as the flattened file fields. Their absence preserves the legacy single-layer shape.
  • Commit and cumulative-diff requests use agentctl's configured repository-qualified ref when present. A caller-supplied branch name cannot override it.
  • The internal agentctl control API adds a per-repository comparison-target update alongside the existing base-branch update. It validates identities and refs, materializes the exact ref, replaces tracker state, and triggers a refresh.
  • The bounded task Git summary adds comparison_unavailable; when true, additions/deletions are not rendered as authoritative task-card statistics.

Failure modes

Scenario Observable behavior
Primary branch or porcelain observation fails The live observation fails and the prior cached snapshot remains available.
Secondary diff enrichment fails The established same-HEAD carry-forward behavior is preserved.
A path has both index and working-tree changes One raw file entry carries both facets. Changes renders one row in Staged and one in Unstaged, while overall file totals count the path once.
A mixed path's layer diff exceeds its cap or the snapshot budget The affected facet retains status metadata and reports the existing truncated or budget_exceeded skip reason; the other facet remains independently usable.
One caller cancels while a shared observation is running That caller returns its context cancellation promptly; other callers remain eligible to receive the shared result.
The tracker stops or the shared deadline expires Underlying work is cancelled and no partial result is published or cached.
A repository has an untracked node_modules tree without an ignore rule Git excludes the tree during untracked enumeration. The snapshot contains no dependency-tree paths or synthetic diffs.
A tracked path exists below node_modules The tracked status query preserves its staged or unstaged change in the snapshot.
Only untracked dependency-tree content changes The monitor fingerprint remains unchanged and does not start a full status refresh.
One repository fails during a multi-repository request Successful repository entries remain available and the failure is reported on its repository entry.
Stored base commit is a strict ancestor of merge-base(HEAD, <resolved integration candidate>) and the target is not a live non-default upstream branch Commit enumeration and cumulative diff use the freshly computed merge-base, not the stale stored SHA; the count reflects only the branch's own commits.
Comparison target is a live non-default upstream branch (e.g. origin/develop) whose merge-base is an ancestor of the integration merge-base The target's own merge-base is preserved; the range is NOT re-anchored to the integration line.
Resolved base branch exists only as a stale local ref (upstream merged/deleted) but an integration origin/* ref is present The origin/* integration ref anchors the range; the stale local ref does not.
No origin/* integration ref and no usable merge-base (unrelated histories) Falls back to the existing stored-base / branch-tip anchor; behavior is unchanged from today.
A linked PR targets a same-named branch in another repository The deterministic comparison remote and exact target ref anchor every comparison-derived surface. origin/<branch> is not consulted.
The PR head repository or branch does not match exactly one task attachment The PR remains associated for Review, but no attachment comparison target changes; the mismatch is logged with task, repository, and branch identity.
The configured comparison remote already exists with another URL Target setup fails closed and reports comparison_remote_conflict; Kandev does not overwrite the remote or fall back to origin.
The configured target branch cannot be fetched File status remains available; comparison-derived data is unavailable with comparison_fetch_failed. Prior or same-named origin statistics are not published as current.
The target needs credentials that are not in the instance environment Git does not read the launcher terminal. The command ends within its deadline and comparison data reports unavailable.
A comparison-target fetch is slow or unreachable Instance readiness and Git-status UI requests continue. The target remains pending or unavailable until the background command ends.
An HTTPS comparison-target fetch fails Kandev does not retry SSH. The selected remotes, upstream, credential scope, and push route remain unchanged.
The backend or executor restarts The persisted target is revalidated and materialized before comparison-derived status becomes authoritative.

Scenarios

  • GIVEN a stale cached snapshot after a commit, WHEN a caller requests fresh=true, THEN the response reflects the live clean tree and a later cached read still returns the prior cached snapshot.
  • GIVEN six simultaneous fresh requests for one repository, WHEN their observations overlap, THEN exactly one underlying status observation runs and all non-cancelled callers receive the same capture timestamp and result.
  • GIVEN simultaneous fresh requests for two repositories, WHEN multi-repository status runs, THEN one observation per repository may run in parallel and each response remains identified with its repository.
  • GIVEN one waiter cancels during a shared observation, WHEN other waiters remain, THEN the cancelled waiter returns promptly and the remaining waiters receive the completed result.
  • GIVEN tracker shutdown or the shared-observation deadline while enrichment is running, WHEN cancellation reaches the observation, THEN filesystem iteration stops and no partial snapshot is cached.
  • GIVEN approximately 15,000 eligible untracked text files outside dependency trees, WHEN fresh status is computed, THEN every path is present, emitted diff content obeys the existing limits, files not enriched after total-budget exhaustion have budget_exceeded, and post-porcelain enrichment remains linear in the number of entries.
  • GIVEN an untracked node_modules tree with thousands of package files and no matching ignore rule, WHEN fresh status is computed, THEN the tree is excluded before per-file status and diff work while an ordinary untracked source file remains visible.
  • GIVEN a tracked file below node_modules, WHEN the file changes, THEN its tracked status remains visible in the workspace snapshot.
  • GIVEN a stable workspace fingerprint, WHEN only untracked files below node_modules change, THEN the next monitor check keeps the same untracked fingerprint and does not start a full status refresh.
  • GIVEN a tracked file with one staged edit and a different unstaged edit, WHEN fresh status is computed, THEN its one path entry contains staged and unstaged facets with independent line totals and HEAD -> index and index -> working tree diffs.
  • GIVEN a newly added staged file that receives another unstaged edit, WHEN Changes opens on desktop or mobile, THEN the same path appears in both Staged and Unstaged, selecting each row shows only that layer, and the overall changed-file count remains one.
  • GIVEN a mixed path visible in both sections, WHEN the user stages or unstages it, THEN one repository-qualified Git mutation runs and the next snapshot shows only the facet that remains.
  • GIVEN one invalid repository in a multi-repository request, WHEN other repositories succeed, THEN the response retains the successful entries and reports the failure only on the invalid repository.
  • GIVEN verification needs writable scratch space, WHEN it selects a location, THEN the location is outside every Git worktree and existing shared caches remain reusable; if a legacy run creates root-level .verify-cache or .tmp, Git status ignores it.
  • GIVEN a session whose stored base_commit_sha is a strict ancestor of merge-base(HEAD, <resolved integration candidate>) (for example a stacked-PR parent branch that has since merged into the integration branch and lost its upstream ref), WHEN the commits panel is requested, THEN the enumerated commit count matches git rev-list --first-parent --count $(git merge-base HEAD <resolved integration candidate>)..HEAD and excludes the commits that already landed on the integration branch.
  • GIVEN a stored base branch that no longer has an upstream ref (the remote branch was deleted) but a local ref of the same name still points at an old branch point, WHEN the commits panel resolves its base, THEN it anchors to the merge-base against the origin/main/origin/master integration ref rather than the stale local ref.
  • GIVEN a session whose stored base_commit_sha equals or is a descendant of the current merge-base against the integration branch, WHEN the commits panel is requested, THEN the stored base is used unchanged and the count is identical to today's behavior.
  • GIVEN a session comparing against a live non-default upstream branch (for example origin/develop) whose merge-base is a strict ancestor of the integration merge-base, WHEN the commits panel resolves its base, THEN it keeps the develop merge-base and does not re-anchor to the integration line, so commits genuinely part of the develop comparison remain visible.
  • GIVEN a worktree with no origin/* integration ref and a HEAD sharing no history with any candidate branch, WHEN the commits panel resolves its base, THEN it falls back to the existing stored-base / branch-tip anchor and does not error.
  • GIVEN a task attached to contributor/widget whose branch targets upstream/widget:main, and the fork's main is hundreds of commits behind, WHEN Kandev associates the matching PR, THEN the task card, commits panel, and cumulative diff compare against the fetched upstream ref and show only the branch contribution.
  • GIVEN the same fork and upstream both have a branch named main, WHEN a repository-qualified target is active, THEN no comparison path resolves that target through origin/main.
  • GIVEN two attachments for the same repository on different branches, WHEN a PR matches one head branch, THEN only that attachment receives the target and the sibling's comparison remains unchanged.
  • GIVEN a historical PR for a previous checkout branch, WHEN provider sync refreshes it, THEN it cannot replace the current branch's comparison target.
  • GIVEN a live cross-fork target, WHEN the backend restarts and the session resumes, THEN Kandev reconstructs the same comparison remote/ref before publishing authoritative totals.
  • GIVEN an explicit target that cannot be fetched, WHEN Changes and task-list surfaces refresh, THEN working-tree files stay visible, comparison-derived values are marked unavailable, and no inflated same-named-origin total is shown.
  • GIVEN an HTTPS comparison target that requires missing credentials, WHEN Kandev starts an instance, THEN the launcher terminal receives no prompt and instance readiness does not wait for the fetch.
  • GIVEN the same missing HTTPS credentials, WHEN target materialization fails, THEN Kandev reports comparison data unavailable and does not try an SSH URL.
  • GIVEN a slow comparison-target fetch, WHEN a lazy repository tracker is requested, THEN the tracker returns immediately with pending comparison state and file status remains usable.
  • GIVEN a provider-derived target, WHEN the user selects a branch from the attachment's Compare against picker, THEN the provider target is removed atomically and the selected local-repository branch becomes authoritative.

Out of scope

  • Replacing existing Git-status routes.
  • Raising or removing existing diff-content limits.
  • Changing multi-repository fan-out behavior.
  • Making fresh reads owners of the polling cache.
  • Replacing Git subprocesses with a native Git implementation.
  • Rewriting the persisted base_commit_sha as part of staleness detection. The read-time correction changes only which base the commits/diff compute against; persistence continues to follow the existing capture and "Compare against" reset paths.
  • Auto-retargeting the session's base_branch when a stacked parent merges. Detecting a stale base and picking a live integration ref is in scope; changing the stored base branch is not.
  • Granting a new Git credential scope for a private comparison repository.
  • Periodically fetching a moving target branch on every status poll. Targets are refreshed on association, retarget, launch/resume, and explicit comparison refresh.
  • Automatically changing the checkout, push remote, upstream, or origin because a comparison target changes.
  • Redesigning the desktop Changes panel or compressing it into the mobile layout.
  • Adding hunk-level staging or changing the existing whole-file stage, unstage, and discard semantics.
  • Defining new behavior for Git's unmerged/conflict porcelain states; those continue to follow their existing handling.
  • Suppressing untracked .next, dist, build, or other generated directories without a Git ignore rule.
  • Adding client-side dependency-tree filtering or changing status payload fields.
  • Retrying an HTTPS comparison target through SSH or another transport.

Implementation plan

See Workspace Git Status Scalability plan, Workspace Git Status Dependency-Tree Exclusion plan, Fork PR Comparison Targets plan, Non-interactive Comparison-target Git plan, and Mixed Staged and Unstaged Changes plan.

The shared Git execution repair extends enforcement beyond comparison commands.