Gatekeeper loop
Read CHARTER.md and agents/gatekeeper-agent.md. Review only the changes
requested by the current task. The owner controls landing and any standing work
queue. This skill creates no additional publishing or messaging authority.
Review the actual result
- Read the diff, the reported failure, and the input the affected tools consumed.
- Confirm that each claimed measurement came from an executed native tool and its real report.
- Compare native numbers with the original spec, PDK, or owner limit. Keep failed and unmeasured work open.
- Inspect the fix for a general cause, unintended scope changes, and loss of actual tool functionality.
- Check relevant tests and public-entry evidence on the current tree. Preserve the Router's bounded task window.
- Hand the reviewed result to the owner or perform the already-authorized landing action through the owner's process.
For IC work, preserve the expert DB and design craft, analog and mixed-signal programs, FPGA steps, and the MCP tools. For benchmarks, preserve Program First, independent blind AI review and backup, clean-room actor environments, and the official scorer. Do not read hidden oracle, harness, or golden contents.
Captured rule: reject checker-derived intent
Owner ruling, 2026-11-13: reject a rule or generator justified by what a benchmark testbench, scorer or golden answer expects. Require the behavioral claim to follow from the supplied INPUT; score improvement alone does not justify it. Keep useful technique, remove oracle authority and benchmark or problem names from author-facing knowledge. Silent or ambiguous input goes to AI authoring with the ambiguity named, without a checker-derived default.
For a generator, ask for a held-out case whose input independently supports the same general technique, and inspect its actual result. Do not pick the case or its expected behavior from hidden answers. Request this evidence within the owner's run scope; a passed-design hold remains binding. Tool-owned regression goldens can test specified tool behavior; they cannot supply a design's missing requirements.
What converging a design teaches goes into the product
Owner, 2026-11-12: the time spent converging a benchmark IC or IP, finding the
real root cause and then its solution, is kept only when it is extracted into
the plugin, the vibeic-eda fork, or an IC expert or PPA skill. This section
is the one statement of that rule; agents/repo-gatekeeper.md,
agents/gatekeeper-agent.md and fork-gatekeeper-loop refer to it.
A fix found while converging a design is not finished until what was learned is in the product, in this order of preference:
- a program in the plugin, when a program can decide it;
- a patch in our fork of the tool, when the tool is wrong or cannot do it;
- a captured rule in the skill the step routes to, when it needs judgement:
the IC expert's skills, or
ppa-optimizefor PPA. A## Captured rulestates the symptom, the cause, the fix and the measured numbers.
benchmark-enhancement-capture says where each kind goes.
A change that makes a row pass, improves a PPA number or removes a wrong
measurement lands only when its hand-in names the capture and its level.
Check that the step's review reaches the rule: the skill is listed for the
step under flow_steps in benchmark/CAPTURE_ROUTING.json, and the rule's
heading appears in review_skills of the step's result.json. When a root
cause is later shown wrong, correct the rule; do not leave it beside a new
one. A tool patch carries what fork-gatekeeper-loop lists.
"Was it captured" is answered with a count from the repository. Landings since a date by the deepest capture they changed, from the repository root:
git log --first-parent --since=2026-11-10 --format=@%h --name-only origin/main \
| awk '/^@/{if(h)print c; h=$0; c="none"; next}
/\/(programs|flow|mcp-eda)\//{c="program"; next}
/\/(skills\/[^/]+\/SKILL\.md|agents\/ic-expert-agent\.md)$/ && c!="program"{c="rule"}
END{print c}' | sort | uniq -c
Captured rules each step's review receives, from the plugin directory:
python3 -c "import json,sys; sys.path.insert(0,'programs'); import flow_engine as f
for i,s in enumerate(json.load(open('flow/steps.json'))['steps'],1):
print(i, sum(len(k['captured_rules']) for k in f.review_skills(s['name'])), s['name'])"
Read a none landing to see whether it fixed a root cause. A tool patch shows
up as its fork pull request, not in this repository.
Landing and reporting
Use ~/claude-leader/land_one.sh according to its actual instructions. Retain
gen_program_inventory.py, nda_diff_scan_check.py, and
tools/release/assign_version.py. Do not introduce another merge queue,
receipt, identity, qualification, or approval layer.
Report the changed behavior, actual native result, test scope, and anything still unmeasured. Report step names and positions from the run's current step table. Keep all run outputs outside the plugin repository. An owner hold stays in force until the owner changes it.
Smallest fix of the original problem (Charter C10, owner 2026-11-14)
Owner: "fixing PR/issue 不要越搞越複雜,用最小的方式、最小的改動去修正原來的問題." Before a fix, state the original problem in one sentence; keep main's existing path and change only what that problem needs. Review asks only: is the original problem fixed, is anything worse than main, and is the fix general and no-cheating? A defect found in behaviour main already has is a separate known gap, written down, never folded into the same PR. After about two SEND BACKs, or when a PR keeps adding new mechanism (parsers, analyses, inventories, readers) to answer counterexamples, stop and restart from main as the smallest fix of the original problem. A smallest fix enters the plugin or a skill only if it is also general and no-cheating (owner 2026-11-14): it holds for any design meeting the same condition, names no benchmark, problem, design or testbench, and takes nothing from a testbench, reference answer, scorer or score. A design-specific fix belongs in that design, not in the product.