reposix-agent-flow

name	reposix-agent-flow
description	Spawn the reposix dark-factory regression test — a fresh subprocess agent given only `reposix init` and a goal must complete clone+grep+edit+commit+push using pure git/POSIX, including the conflict-rebase and blob-limit recovery cycles. The StrongDM/dark-factory regression harness for the v0.9.0 architecture. Invoked from CI release-gate and from local dev (/reposix-agent-flow).
argument-hint	[sim|github|confluence|jira]
allowed-tools	["Bash","Read"]

reposix-agent-flow — dark-factory regression

This skill is the StrongDM / dark-factory regression harness for the v0.9.0 architecture pivot (FUSE → git-native partial clone). It encodes the central thesis from architecture-pivot-summary.md §4 "Agent UX: pure git, zero in-context learning": a fresh subprocess agent — no reposix CLI awareness, no in-context system-prompt instructions about reposix, no MCP tool schemas — must be able to bootstrap a working tree with a single reposix init invocation and then complete its task using only standard git and POSIX tools.

When to use

• CI release gate — wired into .github/workflows/ci.yml as the dark-factory job (sim) and integration-contract-{confluence,github,jira}-v09 jobs (real backends). Phase 36's deletion of the FUSE crate is only safe if this regression keeps passing.
• Local dev — /reposix-agent-flow (or the Skill tool from a coordinating agent) before any change touching crates/reposix-remote/, crates/reposix-cache/, or crates/reposix-cli/src/init.rs.
• Pre-tag verification — python3 quality/runners/run.py --cadence pre-pr (D-CONV-1, 2026-07-04: scripts/green-gauntlet.sh shim deleted, superseded by the quality-pre-pr CI job) runs the sim variant; the real-backend variants run via the gated cargo tests.

What it asserts

The architecture's "pure git, zero in-context learning" claim is the contract. Concretely:

1. reposix init <backend>::<project> <path> produces a real partial-clone working tree: extensions.partialClone=origin, remote.origin.promisor=true, remote.origin.partialclonefilter=blob:none, .git/ is a real git directory.
2. Conflict path teaches the agent. When git push rejects because the backend drifted, the helper's stderr names git pull --rebase as the recovery move — verbatim. An agent that reads stderr and follows the instruction recovers without prompt engineering.
3. Blob-limit path teaches the agent. When command=fetch would materialize more blobs than REPOSIX_BLOB_LIMIT, the helper refuses with a stderr error that names git sparse-checkout as the recovery move.

These are byte-identical regression tests — the teaching strings are the contract, not implementation details.

Invocation

# Default — sim backend, run from anywhere in the workspace.
bash quality/gates/agent-ux/dark-factory.sh sim

# Real-backend variants delegate to the gated cargo tests in
# crates/reposix-cli/tests/agent_flow_real.rs. Each requires the
# corresponding env-var pack (see docs/reference/testing-targets.md):
bash quality/gates/agent-ux/dark-factory.sh github       # → cargo test --ignored dark_factory_real_github
bash quality/gates/agent-ux/dark-factory.sh confluence   # → cargo test --ignored dark_factory_real_confluence
bash quality/gates/agent-ux/dark-factory.sh jira         # → cargo test --ignored dark_factory_real_jira

When this skill is invoked from the Skill tool with no argument, default to sim — it's the only variant that runs without external credentials and is the canonical regression target per project CLAUDE.md OP-1.

Canonical test targets (real-backend variants)

Per project CLAUDE.md OP-6 and docs/reference/testing-targets.md:

• Confluence space "TokenWorld" (reuben-john.atlassian.net) — owner- sanctioned scratchpad. "Go crazy, it's safe."
• GitHub repo reubenjohn/reposix issues — ours; safe to create/close issues during tests. Cleanup is automatic via gh issue close.
• JIRA project TEST — default key, overridable via JIRA_TEST_PROJECT or REPOSIX_JIRA_PROJECT.

The simulator (reposix-sim) is always the default and is the only variant that runs in autonomous mode without explicit credentials.

How to interpret failures

If dark-factory.sh sim fails:

• reposix init exit non-zero — crates/reposix-cli/src/init.rs regression. Check extensions.partialClone config writes.
• extensions.partialClone != origin — git config not applied.
• BLOB_LIMIT teaching string regressed — someone changed the error message in crates/reposix-remote/src/stateless_connect.rs. The literal string git sparse-checkout is the contract; any rewrite must preserve it.
• conflict-rebase teaching string regressed — someone changed the message in crates/reposix-remote/src/main.rs. The literal git pull --rebase must remain.

References

• architecture-pivot-summary.md §4 "Agent UX: pure git, zero in-context learning" — the central design claim this skill regression-tests.
• quality/gates/agent-ux/dark-factory.sh — the underlying script.
• crates/reposix-cli/tests/agent_flow.rs — the cargo test variant (regression-protects the teaching strings via file-content asserts).
• crates/reposix-cli/tests/agent_flow_real.rs — real-backend gated tests.
• docs/reference/testing-targets.md — env-var setup, cleanup procedure, owner permission statement.
• .github/workflows/ci.yml — dark-factory job + three integration-contract-*-v09 jobs that wire this skill into CI.