Resume Work Lanes

Resume Work Lanes

Rebuild the state of interrupted multi-agent work from live evidence. Do not assume that a transcript claim, old plan, branch name, or existing worktree proves current status.

1. Choose the mode

• Orient: inventory lanes and produce the recovery plan. Use by default.
• Resume: orient first, then restart approved lanes. Enter this mode only when the user explicitly asks to resume, start, or dispatch the lanes.

Do not mutate repositories, switch branches, clean worktrees, or start agents during orientation.

2. Start with Cortex

Read references/cortex-evidence.md.

If the cortex CLI is installed, run the bundled collector:

lane_tmp_dir="$(mktemp -d)"
python3 <skill-dir>/scripts/collect_lane_evidence.py \
  --workspace "${WORKSPACE_ROOT:-$HOME/workspace}" \
  --since-hours 72 \
  --source auto \
  --progress \
  --output "$lane_tmp_dir/lane-evidence.json"

The artifact is created atomically and uses mode 0600 on POSIX. On Windows it inherits the ACL of the output directory, so use a private user-owned directory. Use --stdout only when the output channel is private.

If the CLI is absent, search the live tool catalog for Cortex before falling back:

1. Call Cortex MCP sessions, list_ai_projects, usage_blocks, search_sessions, and project_context using the discovered schema.
2. Run the collector with --source git-only to obtain the live Git/worktree inventory without broadly parsing raw transcripts.
3. Record the source as cortex_mcp in the report.

Only use the collector's automatic raw-file fallback when neither Cortex transport is usable.

Use a longer window when the interruption is older. The collector emits:

• Cortex watcher/index health and any freshness warnings;
• recent Cortex projects, usage blocks, session-search results, project context, canonical transcript paths, and bounded excerpts;
• every discoverable Git worktree with branch, HEAD, upstream, dirty state, unique commits, and changed files;
• correlations between Cortex sessions and worktrees;
• collection errors instead of silently dropping uncertain evidence.

Inspect source_used, coverage, and errors before proceeding:

• cortex: normal path. Use Cortex results as the discovery layer.
• cortex_degraded_with_raw_fallback: Cortex health or an essential query failed, and bounded raw evidence was merged. State every failure and truncation.
• raw_files_fallback: Cortex was unavailable and bounded raw evidence was used.
• git_only: session evidence must come from Cortex MCP.

Do not silently prefer raw transcript scans when Cortex is usable. Do not run cortex sessions index during orientation because it mutates Cortex's index; use the raw fallback or ask before repairing the index.

The collector is read-only except for its requested output file. Treat its JSON as triage evidence, not the final answer.

If coverage.session_inventory_status is incomplete, do not present the lane inventory as exhaustive. Prefer Cortex MCP sessions for a time-bounded inventory; otherwise rerun the CLI collector with a larger --cortex-project-limit, --cortex-context-limit, or --session-index-limit as indicated by the exact limitation flags. Stay within the overall deadline and report anything still omitted.

3. Identify candidate lanes

Include a lane when one or more of these are observed:

• a recent session cwd maps to a worktree or repository;
• a recent plan or agent event names a distinct task;
• a non-primary worktree is dirty, ahead, unmerged, or attached to an open PR;
• a session ended with unfinished plan items, a blocking error, a pending tool call, or a handoff;
• a completed implementation still lacks verification, push, review, merge, deploy, or cleanup.

Do not treat these as active lanes without more evidence:

• protected long-lived worktrees documented by repository policy;
• clean primary branches with no incomplete recent session;
• merged or deleted branches whose cleanup is already complete;
• old session files that have been superseded by a newer session on the same goal.

Separate one session containing several independent agent tasks into several lanes. Combine multiple sessions only when they clearly continue the same goal in the same implementation line.

The collector keeps session_candidate and worktree_snapshot records separate. A mentioned worktree is a dependency, not proof that the session owns that lane. Treat inventory_only_unverified and text search_hits as leads that cannot create an active lane without stronger evidence.

4. Verify each lane

For every candidate lane:

1. Read the repository's CLAUDE.md before interpreting branch or worktree state.
2. Re-run live Git status in the exact worktree. Record branch, HEAD, upstream, ahead/behind, dirty files, recent commits, and diff against the intended base. The collector uses local refs and does not fetch. Treat local_ref_only_no_fetch merge results as provisional until forge verification.
3. Use Cortex search_sessions/sessions search and project_context/sessions context to recover:
   • the user goal and scope changes;
   • the last explicit plan and its statuses;
   • subagent task assignments and returned results;
   • files claimed changed and commands claimed run;
   • blockers, failures, open questions, and the last promised next action.
4. Read a raw transcript only when Cortex returns a canonical transcript path and the indexed snippets do not establish an exact fact. Do not rescan every transcript by default.
5. Verify claims against files, diffs, commits, tests, and generated artifacts. A transcript saying "done" is not proof.
6. Check repository-native trackers and plans when present. Use bd for Beads repositories and inspect docs/plans, .claude/current-plan, or equivalent live artifacts.
7. Check the forge when configured: open PR, review state, check runs, merge state, and whether the branch still exists remotely.
8. Record drift between the interrupted session and now. Never paper over moved HEADs, changed bases, new commits, resolved blockers, or deleted worktrees.

Bound expensive reads. Start with transcript tails and targeted searches; expand to the full transcript only when required to establish a fact.

5. Classify lane state

Use exactly one state:

• active_incomplete: useful work exists and concrete implementation remains.
• blocked: work cannot proceed until a named dependency, decision, permission, or external state changes.
• completed_unlanded: implementation and required verification are complete, but push/review/merge/deploy remains.
• landed_needs_closeout: work landed, but sync, cleanup, deployment verification, tracker updates, or handoff remains.
• completed: the goal and required closeout are proven complete.
• protected_long_lived: repository policy says the branch/worktree intentionally persists.
• unknown: evidence is insufficient or contradictory.

State confidence as high, medium, or low and cite the evidence that controls the classification.

6. Build the restart plan

Read references/report-template.md and use its report shape.

Order lanes by:

1. blockers that unlock other lanes;
2. nearly complete landing/closeout work;
3. foundational shared changes;
4. independent implementation lanes;
5. cleanup only after dependent work is safe.

Call out overlapping files, shared migrations, branch ancestry, deploy dependencies, and resource contention. Mark lanes as parallel only when they can run independently.

For every non-complete lane, produce a restart packet containing:

• exact worktree and branch;
• verified objective and current state;
• completed implementation with evidence;
• remaining checklist;
• first command or file inspection;
• required tests and completion gate;
• dependencies and forbidden scope;
• transcript, plan, bead, PR, and CI references.

7. Resume only with authority

If the user explicitly requested resume mode:

1. Present the inventory and restart order first.
2. Preserve dirty and unmerged worktrees.
3. Reuse the existing worktree when safe; create a new one only when the prior path is missing or unsafe.
4. Start replacement agents only when the environment permits delegation and the user's request authorizes it. Old in-memory agents cannot be resumed.
5. Give each agent one bounded lane packet. Do not leak conclusions from unrelated lanes.
6. Track every restarted lane and report whether it is running, blocked, or completed.

If authorization is orientation-only, end by asking which lanes to activate. Do not start work implicitly.

Quality rules

• Prefer live repository and forge evidence over transcript recollection.
• Prefer Cortex over broad raw transcript parsing; record Cortex health and the actual evidence source.
• Treat broad Cortex keyword matches as leads, not proof. System prompts and tool documentation can mention agent or plan terms.
• Preserve unrelated dirt and protected worktrees.
• Distinguish planned, attempted, implemented, verified, pushed, merged, deployed, and cleaned up.
• Cite paths, SHAs, PRs, run IDs, plan items, and session IDs.
• Label unknowns; never turn absence of evidence into completion.
• Redact credentials and private message content from reports.
• Keep the executive summary concise while making every restart packet executable.
• Report every failed call, truncated search/context/fallback, omitted project/repository, and unverified inventory-only session.