Last Human Commit 1.5

An automatic engineering workflow for ordinary requests, including rough voice transcripts. Understand the intended result, do the useful work, test it and integrate the shared result.

What changed

Operating instructions, internal planning, new task records and agent/tool messages use English. User answers use the user’s language; original messages, localized fixtures and raw feedback retain their language.

Native hooks load instructions at session start or context restoration and issue time-only 30-minute planning and hourly goal-check reminders. Codex also has an optional client-owned async progress adviser: once per due interval, a fresh read-only native Overseer compares bounded original user input, raw recent actions/results and candidate plan state, then returns one next action. The Lead acts/assigns it or rejects it with a concrete observed fact; advice never blocks delivery. Native hook trust and disablement remain in force. Other clients use the portable method manually; their automatic execution is not claimed. They do not classify conversation text, mutate task contracts, or block completion. Every 30 minutes the Lead updates the task note with roughly 15-minute steps, initial estimate versus actual effort, dependencies and parallel delegation. The Lead owns the goal, applicable checks and delivery; existing explicit contract and review commands remain available without automatic conversation parsing. See the outcome contract.

Independent native review remains bound to the reviewer's current task assignment. When work is genuinely reassigned, use a new native worker session under the existing mandate and assign the existing reviewer again with the fresh outcome ID and full current goal. Old reports remain historical and cannot be reused for the new scope.

Agent Herder relays are marked as agent-relay transport origin; unmarked origin remains unknown. This marker preserves provenance and does not grant authority.

Parallel writers use Git worktrees. Ready work converges into the explicitly permitted published dev history; PR per worker is optional. Merge only with authorization, review and green checks. Before finishing, fetch, merge newer work, retest and push. A Full cycle converges the participants' changes, including nested repositories, into one reachable green history. Foreign active work is preserved.

A blocker delays only the operation that needs it. Finish independent work first. Agent Herder's resume_after can wake the original session after selected sessions finish their native execution. A finished session is not automatically a correct result.

Shared coordination uses one atomic JSON file per message. Project worktrees share <primary-checkout>/.agents/shared-session/chat/; host-wide communication uses ~/.agents/shared-session/chat/. Cross-host communication uses an explicit shared address, not coincidentally identical local paths.

Design uses three passes: working solution, ideal version, then the smallest complete implementation preserving roughly 80% of useful value. These are not three builds or approval rounds; the method is separate from a three-line form.

No mandatory setup interview or initial Overseer. Hourly original-goal direction review and 120-minute workflow improvement are mandatory. Fresh Zero Knowledge groups accept ready major milestones, not every hour. Architecture, decomposition, independent review/testing and learning tools remain available; scheduled inspection is required even when the current plan looks correct. The complete default workflow is in AGENTS.md. Details: docs/V1.4.md in the source repository.

Package and installation

The native Agent Plugin is plugins/last-human-commit, version 1.5.48, with the full optional skill set (validated by scripts/validate.py), instruction-only startup hooks and time-only reminders, plus the Codex-specific async extension. Use the harness's native plugin manager. For clients whose plugin hooks are unavailable or still await trust, the package provides a supported user-instruction bootstrap:

python3 <installed-package>/scripts/install_runtime.py \
  --runtime codex --plugin-root <installed-package> --replace-gsd-entry

--runtime also accepts claude and opencode. Only managed LHC/GSD entry blocks are replaced; unrelated instructions are preserved and backed up. Existing GSD skills are not deleted. Reload/start a new session to load changed instructions.

Source updates: edit the source, run scripts/sync_skills.py, then python3 tests/validate.py in the source repository. The validator runs package/source parity and all executable regressions. Runtime installation and actual use must be verified separately from source tests.

1.4.3 clarification

Define a tangible user/consumer result before code: who needs what, before/after and where to try it. Start bug TDD immediately. Hourly direction reviews remain; Zero Knowledge groups accept major outcome milestones (roughly 3 hours), not every hour or edit. Varied realistic personas leave detailed positive/negative feedback. Preserve raw feedback, fix actual obstacles and keep preferences without automatically expanding scope. See docs/V1.4.3.md in the source repository.

OpenCode loads its normal AGENTS.md and skill discovery, and the native LHC plugin (opencode/lhc-v2.ts, registered by scripts/install_runtime.py into the OpenCode plugin directory) injects the current installed AGENTS.md into the system context, reports the time-only 30/60-minute reminders at tool and idle boundaries, and re-injects the AGENTS.md and bounded current-handoff.md pointers after compaction. It does not review, gate, approve or evaluate a diff. Hermes retains only explicit role-selection middleware, without a recurring lifecycle observer. Codex/Claude/ZCode load concise instructions and time-only 30/60-minute reminders. The reminders run at the next tool boundary, not through a background scheduler. Native manager installation does not prove that every existing chat has reloaded its registered hooks; verify the loaded version when diagnosing an old chat.