manual_slop/docs/guide_tier2_autonomous.md

Tier 2 Autonomous Sandbox

Why this exists

When you run Tier 2 in the main repo, every edit and every bash call prompts you for approval (permission: ask). For well-regularized tracks (TDD red/green with atomic per-task commits), this is noise. This track adds an autonomous mode in a sibling clone where Tier 2 runs unattended, with a 3-layer enforcement stack to keep it contained.

One-time bootstrap

cd C:\projects\manual_slop
pwsh -File scripts\tier2\setup_tier2_clone.ps1 -WhatIf   # dry run first
pwsh -File scripts\tier2\setup_tier2_clone.ps1            # actual bootstrap

The bootstrap:

1. Clones the main repo to C:\projects\manual_slop_tier2\
2. Sets origin = C:\projects\manual_slop (local path; no remote)
3. Copies the agent, slash command, and opencode.json templates to the clone
4. Installs the git hooks (pre-push refuses all pushes; post-checkout logs checkouts)
5. Creates a "Tier 2 (Sandboxed)" desktop shortcut

As of 2026-06-18: the bootstrap no longer creates any directory on AppData. Tier 2 state and failure reports live at tests/artifacts/tier2_state/<track>/state.json and tests/artifacts/tier2_failures/<track>_<ts>.md (project-relative; inside the project tree under the already-gitignored tests/artifacts/). The user directive is "NEVER USE APPDATA" — enforced by the OpenCode *AppData\\* bash deny rule.

Per-track invocation

1. Double-click the "Tier 2 (Sandboxed)" desktop shortcut (or run pwsh -File C:\projects\manual_slop\scripts\tier2\run_tier2_sandboxed.ps1 manually)
2. In the OpenCode session, type:

   /tier-2-auto-execute <track-name>
   

   Examples:
   • /tier-2-auto-execute result_migration_review_pass
   • /tier-2-auto-execute data_structure_strengthening_20260606 --resume
   • /tier-2-auto-execute rag_test_failures_20260615 --toast
3. Tier 2 runs the track autonomously, commits per task, monitors failcount
4. On success: prints a summary
5. On give-up: writes a failure report and prints the path

Review and merge

After Tier 2 finishes (success or give-up):

1. cd C:\projects\manual_slop (back to main)
2. git fetch C:/projects/manual_slop_tier2 tier2/<track-name>
3. Review the diff with Tier 1 (interactive)
4. On approval: git merge --no-ff tier2/<track-name> to main

The 4 hard bans (enforced at 3 layers)

Ban	Layer 1 (OpenCode)	Layer 2 (OS)	Layer 3 (git hook)
git push* (any push)	permission.bash deny rule	n/a	pre-push hook refuses all pushes
git checkout* (any form)	permission.bash deny rule	n/a	post-checkout hook logs the checkout
git restore* (any form)	permission.bash deny rule	n/a	n/a
git reset* (any form)	permission.bash deny rule	n/a	n/a
File access outside Tier 2 clone (AppData, Temp, Documents, etc. all denied)	permission.read/write path allowlist + *AppData\\* bash deny	Windows ACL	n/a

The failcount threshold

Tier 2 gives up if ANY of these hit:

• 3 consecutive red-phase failures (the test doesn't fail when it should)
• 3 consecutive green-phase failures (the implementation doesn't make the test pass)
• 30 minutes with no progress (no commit, no green test)

Override via scripts/tier2/failcount.toml.

The failure report

Written to tests/artifacts/tier2_failures/<track>_<timestamp>.md (project-relative; inside tests/artifacts/ which is gitignored) with 7 sections:

1. Header (track, branch, started, stopped, duration, give-up signal)
2. Tasks completed
3. Current task (where it stopped)
4. Last 3 failures
5. Failcount state
6. Git state (git log tier2/<track> ^origin/master)
7. Recommendation (heuristic-based)

A .STOPPED flag file is created alongside the report. The main repo can check for it on next Tier 1 session start (an opt-in banner).

Conventions (added 2026-06-17)

These are enforced by the Tier 2 agent prompt. The agent MUST follow them — they're not optional.

• Test runner: Tier 2 always uses uv run python scripts/run_tests_batched.py. Never uv run pytest directly. The batched runner provides tier-based filtering, parallelization (xdist), and a summary table that direct pytest doesn't.
• Default branch: this repo uses master (not main). When fetching or branching, use origin/master. Tier 2 may otherwise get confused by the missing main reference.
• Line endings: Tier 2 preserves existing line endings on edit. This repo has a mix of CRLF and LF; standardizing to repo-wide LF is a future track. For now, do not normalize.
• Throw-away scripts: Tier 2 writes its working scripts to scripts/tier2/artifacts/<track-name>/, NOT the base scripts/tier2/ directory. The base directory is reserved for production code. Throw-away scripts are kept for archival but isolated in a track-specific subdir.
• End-of-track report: at the end of every track, Tier 2 writes docs/reports/TRACK_COMPLETION_<track-name>.md (follow the precedent set by TRACK_COMPLETION_tier2_autonomous_sandbox_20260616.md) and updates conductor/tracks/<track-name>/state.toml to status = "completed". The user reads this report to decide merge.
• Run-time expectation: tracks are expected to take 1-4 hours. If the model reports it is running out of context, Tier 2 notes progress to disk and continues. The user expects autonomous runs to complete without manual "press continue" intervention.

Verify the sandbox (manual checklist)

After bootstrap, run these inside the Tier 2 sandboxed OpenCode session to verify the bans are enforced:

• Try git restore tests/test_failcount.py — should print "denied"
• Try git push origin master — should print "denied" (or the pre-push hook fires)
• Try git checkout -- src/foo.py — should print "denied"
• Try git reset --hard HEAD~1 — should print "denied"
• Try to read C:\Users\Ed\Documents\test.txt (from a Python subprocess) — should print "ACCESS_DENIED"

And verify allowed operations work:

• git status — works
• git switch -c test-branch — works
• Edit a file in the Tier 2 clone — works
• git add <file> && git commit -m "test" — works

Troubleshooting

• "Tier 2 (Sandboxed) shortcut doesn't work": check that pwsh.exe is on the PATH (where.exe pwsh).
• "Permission denied" on file access inside the sandbox: the Windows ACL may be too restrictive. Re-run the bootstrap (setup_tier2_clone.ps1 is idempotent).
• "Failcount state not found": the tests/artifacts/tier2_state/<track>/ dir may be missing. The failcount module creates it on first save; check that the Tier 2 clone's project root is correct.
• "Pre-push hook not firing": check that .git/hooks/pre-push is executable. On Windows, Git Bash runs the hook; check git config core.hooksPath if you have a custom hooks dir.
• "Tier 2 keeps giving up at 30 min": increase no_progress_minutes in scripts/tier2/failcount.toml.
• "Tier 2 ran out of context": the model stopped mid-track. The user (interactive Tier 1) should cd to the Tier 2 clone, inspect tests/artifacts/tier2_state/<track>/state.json for the last completed task, and re-invoke with /tier-2-auto-execute <track-name> --resume to continue. The state file persists across runs.