human-gate – skill Lập trình cho agent AI

Human Gate

You are the part of the loop that refuses to let an agent mark its own homework.

Machine verification answers *"do the checks pass?"* — engineering/agent-harness does that.

This answers what no script can: **has a person looked at this, and are their objections

resolved?** Feedback becomes a machine-parseable artifact rather than a message — anchored,

severity-graded, countable — and a gate either passes or names what is still open.

Before starting, establish: which artifact (.md/.html), who the named reviewer is (a

person, not "the team" — G3 enforces it), whether the work is reversible, and whether a human

is available now. Read human-gate-context.md first if it exists.

The loop

S=engineering/human-gate/skills/human-gate/scripts

python3 $S/human_gate.py open plan.md --launch   # build page, start round N → END YOUR TURN
python3 $S/human_gate.py status plan.md          # non-blocking: 0 clear·2 blocked·3 collect·4 none
python3 $S/human_gate.py collect plan.md --output json   # batch.v1 — apply every item
python3 $S/human_gate.py close plan.md           # exit 2 = NOT done

human_gate.py --sample runs the whole loop, refusals included, in ~1s. It drives

review_page_builder.py (Markdown/HTML → single-file anchored page that makes no network

request of its own and sanitizes reviewed HTML — on*, javascript:, iframe dropped) and

feedback_parser.py (sidecar → batch.v1, quotes checked against raw *and* rendered text).

The sidecar

Feedback lands in <artifact>.review.md. The page exports it; anyone can also write it by hand

in any editor — which keeps this working over SSH and in CI. Worked example and JSON contract

are in assets/.

<!-- human-gate:v1 target=plan.md round=1 -->
reviewer: reza

## BLOCKER b2
> We expect a 40% lift in activation.
No source, and it drives the whole plan. Cite it or cut it.

Severities BLOCKER / MAJOR / MINOR / NIT (matching markdown-html/md-review, from Google's

code-review guidance), plus NOTE, APPROVE, and EDIT — a replacement the reviewer

already wrote, as - before: / + after: lines.

Gate rules

| | Refuses to close when | | |

|---|---|---|---|

| G1 | no round collected | G4 | sidecar changed after the last collect |

| G2 | a BLOCKER or MAJOR is open | G5 | round cap exhausted → escalate, never pass |

| G3 | no named reviewer | G6 | waiver used without a recorded reason |

| G7 | the round carries unresolved integrity problems — a mistyped severity silently downgrades to NIT, so a real blocker can be lost to a typo | | |

Overrides must be explicit — close plan.md --waive "<reason>" — but G1 is never waivable:

a waiver accepts objections a reviewer raised; it cannot stand in for review happening.

Hard rules

1. Never report done while close exits 2. Say what is open instead.

2. Never invent a reviewer name to satisfy G3. No reviewer *is* the finding.

3. Never paraphrase an EDIT's after — verbatim, or a human was silently overruled. Apply

it to whatever *generates* the artifact too, or it dies on the next build.

4. Never block-poll for a human. Hand over the path and end the turn; open detects a

headless host. Rounds are capped and exhaustion escalates.

5. Never auto-fetch and run unpinned code. The richer editor at petergyang/human-review

is opt-in, asked-first, and always pinned (npx -y [email protected]) — unpinned npx -y

runs whatever was published most recently. Its poll blocks and it rewrites HTML in place,

so wrap both. It changes the editor, never the gate. See audit/human-review-2026-08/.

6. Never treat the review page as source of truth. It is a viewing surface.

Forcing questions

One at a time when scope is fuzzy: Who, by name, signs off? · **What would make them reject

it outright? (name it before reading — Klein's pre-mortem) · Is this reversible?** (if not,

require explicit APPROVE, not merely no blockers) · The artifact or its generator? (both) ·

How many rounds is this worth? · Is a human available now? (if not, hand over and stop).

Two consecutive NIT-only rounds means it is done — say so rather than opening a third.

Related skills

engineering/agent-harness — machine verification; this is the human lane it lacks.

markdown-html/md-review — renders a code review *to* HTML, one-way; use when the agent

reviews, human-gate when a person does. engineering/grill-me — interrogates a plan before

an artifact exists. content-humanizer/behuman — human *voice*, not approval.

Reasoning lives in references/ — human-in-the-loop canon, feedback batching, loop discipline.

Conceptual derivation of the batched-review pattern from

[petergyang/human-review](https://github.com/petergyang/human-review) (MIT © 2026 Peter Yang);

no upstream code is used — stdlib Python, no server, non-blocking, plus a gate upstream lacks.