Targeted worktree get and finalize commands hang on a clean managed worktree

[chubes4]

Problem

Targeted DMC lifecycle commands hung during a Homeboy PR finalization handoff on a clean, pushed managed worktree. This prevented Homeboy from reading and recording DMC finalizer metadata after it opened the PR.

Observed with Data Machine Code 0.53.6 on macOS/WordPress Studio.

Reproduction

Managed worktree:

• Handle: wp-codebox@perf-2135-canonical-restore-pack
• Path: /Users/chubes/Developer/wp-codebox@perf-2135-canonical-restore-pack
• Branch: perf/2135-canonical-restore-pack
• State: clean and tracking origin/perf/2135-canonical-restore-pack
• PR: Add content-addressed Cloudflare canonical restore packs Automattic/wp-codebox#2142

From the Studio site root, run:

studio wp datamachine-code workspace worktree get wp-codebox@perf-2135-canonical-restore-pack --format=json --path=/Users/chubes/Studio/intelligence-chubes4

Observed: no output; caller timeout killed it after 120 seconds.

Then run:

studio wp datamachine-code workspace worktree finalize wp-codebox@perf-2135-canonical-restore-pack --pr=https://github.com/Automattic/wp-codebox/pull/2142 --path=/Users/chubes/Studio/intelligence-chubes4

Observed: no output; caller timeout killed it after 300 seconds.

The worktree remained clean, and direct git status --short --branch returned immediately. GitHub confirmed the PR was open at exact head 0cae6497c10fc4c7daad95ed1ef20d2a32b43e8f.

Expected

• worktree get should return the one requested row within a bounded deadline.
• worktree finalize should either persist and return lifecycle metadata or fail with a bounded, actionable error.
• A timeout/interruption must expose whether metadata committed before the caller was terminated.
• Homeboy should be able to complete DMC bookkeeping after PR publication without an unbounded CLI process.

Source evidence

• WorkspaceCommand::WORKTREE_OPERATIONS maps get to datamachine-code/workspace-worktree-list.
• WorkspaceCommand::worktree() forces include_status=true for get.
• WorkspaceWorktreeLifecycle::worktree_list() calls count_unpushed_commits($wt_path) without a timeout for that targeted row.
• datamachine_code_is_targeted_workspace_read_cli_request() recognizes only workspace show, not workspace worktree get, so the targeted read does not receive the side-effect-free bootstrap treatment.
• worktree_finalize() bounds its dirty Git probe, then writes lifecycle metadata and upserts inventory; the overall CLI call still had no externally observable completion within five minutes.

Relevant installed-source locations:

• data-machine-code.php:55-63
• inc/Cli/Commands/WorkspaceCommand.php:45-49,4146-4154
• inc/Workspace/WorkspaceWorktreeLifecycle.php:528-586,822-932

Acceptance criteria

• Add a true targeted handle lookup that avoids scanning unrelated workspace state.
• Apply explicit deadlines to every Git/database/runtime step reachable from targeted get and finalize.
• Emit bounded progress or phase diagnostics identifying the blocking step.
• Add regression coverage for a clean pushed worktree with a PR URL.
• Prove both commands complete under a deterministic wall-clock bound in a large-workspace fixture.

Context

Homeboy cook: agent-task-040917dd-8e78-4ef4-abf5-6e052784e1de. All code/test gates passed and PR creation succeeded; only the DMC metadata handoff remained unverified.

AI assistance

OpenAI openai/gpt-5.6-sol via OpenCode gathered the command evidence, inspected the installed DMC source, checked for duplicate issues, and drafted this report. Chris Huber reviewed and filed it.

[chubes4]

Reopening with a fresh reproduction on 2026-07-28 after the earlier cleanup appeared to restore service.

Current managed handle: homeboy@fix-10588-detached-cook-recovery. The worktree exists, is clean, and was created successfully through DMC. Homeboy continuation then failed before promotion with:

worktree provider `dmc` resolve command timed out after 10000 ms

Direct targeted reproduction:

studio wp datamachine-code workspace worktree get homeboy@fix-10588-detached-cook-recovery --format=json

Observed: no output; caller killed it after 180 seconds. A subsequent studio wp plugin get data-machine-code --format=json also produced no output before a 120-second timeout, and studio wp datamachine-code workspace show data-machine-code timed out after 180 seconds. This indicates the current failure may be broader WordPress/DMC process or lock starvation rather than only one worktree scan.

Impact chain:

• Blocks Engine Fix bulk abandoned worktree cleanup limits #700 candidate is stranded behind Homeboy #10588.
• Homeboy #10588 repair candidate exists but its legal cook-continue path cannot resolve the owning DMC worktree within Homeboy’s 10-second bound.
• No provider budget was consumed by the failed continuation.

Earlier workspace hygiene reported 4,535 DB lock rows, 50 stale filesystem locks, 358 worktrees, and no active locks; after user cleanup, one worktree add succeeded, but targeted reads later regressed to unbounded hangs.

Needed evidence/fix: identify the blocked WP/DMC phase and expose bounded phase diagnostics; targeted get must stay independent of unrelated inventory/lock volume, and a stale/crashed request must not make ordinary WP CLI reads hang.

AI assistance: OpenAI openai/gpt-5.6-sol via OpenCode/Homeboy reproduced the nested recovery failure and captured the timeout chain. Chris Huber reviewed this update.

[chubes4]

Fresh root-cause evidence from the 2026-07-28 recurrence:

A live sample of the stalled Studio WP-CLI cron process showed PHP blocked in php_exec() reading from a child process, with the remainder of the loop sleeping/polling. The direct child was:

sh -c du -sk /Users/chubes/Developer/wp-codebox@investigate-1916-latest-diagnostics/node_modules 2>/dev/null

The owning source is inc/Workspace/WorkspaceHygieneReport.php:531-546. directory_size_bytes_best_effort() invokes synchronous exec("du -sk ...") with no deadline, cancellation, or per-entry budget. It runs from the scheduled workspace_hygiene_report task registered in data-machine-code.php:404,455-458.

Observed impact:

• The cron WP-CLI process remained alive for more than eight minutes while scanning one worktree dependency tree.
• Studio’s full-wp-load.php installation probe took more than five minutes.
• The four HTTP workers progressed from roughly 6-second requests to 60+ second occupancy, and /wp-json/ returned zero bytes before a 20-second client timeout.
• Targeted DMC provider resolution exceeded Homeboy’s 10-second boundary.
• SQLite PRAGMA quick_check returned ok; no Markdown Database Integration frame appeared in the sampled blocker.

Acceptance addition: scheduled hygiene size attribution must use a supervised process with an explicit per-entry and total deadline, terminate the full du process tree on timeout, record the skipped path/reason, and yield bounded partial evidence instead of blocking cron or ordinary control-plane operations.

[chubes4]

Fresh recurrence on 2026-07-28 while proving Extra-Chill/homeboy#10643 for Static Site Importer fixture 37.

The targeted get repair from #965 is present, but Homeboy path-to-worktree resolution still invokes the configured provider list command:

studio wp datamachine-code workspace worktree list --with-status --format=json --path=/Users/chubes/Studio/intelligence-chubes4

Observed on a workspace with 387 worktree-like directories:

• Initial direct list completed in 18.7s and emitted roughly 46,000 lines of valid JSON.
• Subsequent list exceeded 180s under load.
• A final candidate Homeboy run gave DMC a declared 300,000ms provider budget and 64 MiB output budget. DMC still did not complete; Homeboy terminated the full process tree at 300,287ms.
• Before the output budget was configurable, Homeboy truncated this inventory at 64 KiB and then surfaced a misleading invalid-JSON error. Configure worktree provider lookup bounds homeboy#10643 fixes those caller-side timeout/output contracts and preserves bounded termination.

Impact: homeboy bench --placement local --path=<managed-worktree> cannot reach WP Codebox execution because path resolution enumerates every worktree with status probes. The SSI evidence-contract proof remains blocked before workload start.

Needed owning-layer repair: provide a bounded targeted path-to-handle resolver, or a compact bounded list projection that returns only Homeboy’s required {handle,path,branch,safety} fields without exhaustive metadata/status work. Path resolution should not scale with unrelated worktree count.

AI assistance: OpenAI openai/gpt-5.6-sol via OpenCode reproduced the failure, implemented the bounded Homeboy caller contract, and captured the five-minute termination evidence. Chris Huber reviewed this update.