fix: resume uses snapshot directory instead of caller's cwd

[bloodcarter]

Summary

• hcom r <name> now starts the agent in the directory where it was originally launched, not where the resume command is run
• Fixes database corruption when resuming from a different directory
• Adds --dir flag for explicit directory override on resume/fork

The Bug

When resuming an agent with hcom r <name> from a different directory than the original launch, two things go wrong:

1. Wrong working directory: The resumed agent starts in the caller's cwd instead of the directory stored in the stopped snapshot. This breaks relative paths, project-specific configs, and CLAUDE.md / .claude/settings.json resolution.
2. Database corruption: initialize_instance_in_position_file() independently reads current_dir() and writes it to the directory field in the DB. If the agent is later stopped and resumed again, the corrupted directory propagates through snapshots.

Reproduction

cd /project/frontend && hcom 1 claude --go   # launches "luna" in /project/frontend
hcom kill luna
cd /project/backend && hcom r luna            # BUG: luna starts in /project/backend

The Fix

1. resume.rs — Use snapshot directory for resume

do_resume() now extracts the directory field from the stopped snapshot and computes effective_cwd with this priority:

Priority	Condition	Directory used
1	--dir /path flag	The specified path (validated + canonicalized)
2	Fork mode (hcom f)	Caller's current_dir() (fork = fresh start)
3	Resume with valid snapshot dir	Snapshot directory (continue where you left off)
4	Resume, snapshot dir deleted	Caller's current_dir() with stderr warning

2. instances.rs — Eliminate stale directory window

initialize_instance_in_position_file() now accepts an optional cwd_override parameter. When provided, it uses the override instead of calling current_dir() independently. This eliminates the window where the DB records the wrong directory during pre-registration.

3. launcher.rs — Pass correct cwd to pre-registration

The launch() function now passes working_dir (from LaunchParams.cwd) to initialize_instance_in_position_file(), ensuring the DB directory is correct from the moment the instance is created.

4. New --dir flag

Resume and fork now support --dir <path> (and --dir=<path>) to explicitly override the working directory:

hcom r luna --dir /project/frontend    # resume in specific directory
hcom f luna --dir /project/new-feature # fork into specific directory

The --dir path is validated (is_dir() check) and canonicalized to an absolute path.

Files Changed

File	Change
src/commands/resume.rs	Extract directory from snapshot, compute effective_cwd, add --dir flag
src/instances.rs	Add cwd_override: Option<&str> parameter
src/launcher.rs	Pass working_dir as cwd_override
src/commands/start.rs	Pass None for new parameter (2 call sites, no behavior change)

Testing

All scenarios verified with isolated HCOM_DIR:

Scenario	Result
Resume from wrong directory → uses snapshot dir	PASS
--dir /tmp override	PASS
--dir /nonexistent → error message	PASS
Resume when original dir deleted → warns, uses cwd	PASS
Fork from different dir → uses current dir	PASS
Fork directory vs original directory are independent	PASS

Works for all tool types (Claude, Codex, Gemini, OpenCode) — the directory logic is in the common path before tool-specific branching.

🤖 Generated with Claude Code