Repository navigation
Replace the detailed diagram with the flow one - #9
Merged
Merged
Conversation
The pipeline diagram had grown to seven steps carrying every weight, every threshold and every rubric level. It documented the sample accurately and read like a specification. The flow diagram says the same shape in five boxes and one branch: a pull request, pre-triage on plain rules with no model, a fork where any of draft, a failing check or a running check stops and reports that state, otherwise one Jev call of eleven questions, the nine becoming an effort and consequence pair, and the pair picking one route. The route table now says what the repo owner does rather than what evidence is sufficient, because that is the sentence a reader wants. It ends with two worked examples, so a reader can see what a skill running this would actually report: apache/airflow #73729 "Fix clearing with upstream and downstream..." bugfix, check correctness first, 2 files +51/-2 security surface 0.06, blast radius 1.1/2, has tests 0.98 -> (0.27, 0.15) -> tests-are-enough mcp-gateway-registry #1693 "feat(security): Asymmetric (ES256) JWT signing" feature, check security first, 51 files +3052/-626 security surface 1.00, blast radius 2.0/2, has tests 0.96 -> (0.74, 1.00) -> human-plus-author Every figure comes from a committed report. The second example is from the gateway run because the airflow report no longer holds a human-plus-author: the pull request that did is now ci-failing, so pre-triage takes it out before Jev sees it. Both rows name their repository rather than mixing quietly. pipeline.html and pipeline.png are deleted, render.py goes back to one target, and the AGENTS.md layout line follows.
mean and max are the reason the pair exists, and they were reading as ordinary words in the same grey as the rest of the line. Each now carries its own colour: teal for the mean, because averaging is the calm operation, and red for the max, because it is the alarm. A reader scanning the diagram now sees that the two halves do different arithmetic before reading either label.
… pass Accuracy first, because the prose had drifted from the code. The worked example cited airflow #73698, a pull request the committed report no longer holds: the dataset was re-fetched, the eighteen most recent open ones changed, and it dropped out. None of its nine rows could be checked against anything in the repo. Rebuilt on #73741, which is in the report and teaches the point better: "Improve DAG tag length validation error", two files, eleven added lines, and it still needs a person because a length check on user input is input validation and security_surface came back 0.97. Averaging the nine puts it at 0.40, mid-queue, where nobody looks twice. That rebuild exposed a second gap. The nine rounded rows sum to 0.3969 while the report records 0.3970, because the report rounds the total rather than adding rounded rows. The table shows 0.3969 and one clause explains it, so a reader adding the column gets what is printed. go/README.md had no mention of pre-triage, though the binary does it. It now carries the states table, the shared-failure rule, the ten-of-eighteen figure, state_counts among the outputs, and a corrected drift line: effort moves about 0.03 and consequence about 0.07, with the 16-of-18 parity result. Smaller fixes: the sample README said twenty-six pull requests want a token, left from the gateway example, and one pull request now costs three GitHub calls rather than two since the checks call landed. It pointed at -fail-on-tier where the route gate is -fail-on-route. Twenty claims verified against questions.yml, the source and the committed report: route floors, effort raises, the size floor, the deadband, eleven questions with nine weighted and three inverted, the four consequence markers, the budgets, the state and route counts, and the parity result. A new section explains why the shape is worth it, framed around who makes the decision rather than around tokens saved. Pre-triage removes ten of eighteen before any model runs, and no model could do that, because no model knows whether CI passed. Deciding what a human sees is itself a model call and should be the cheap one: eleven judgements and two numbers, about 3,000 input tokens whoever does it, $0.00013 at TypeSafe's published price. A reader divides their own model's input price by 0.042 to see the multiple, which keeps a competitor's pricing out of a file that would rot around it. The expensive model then reviews the one pull request in eighteen where a review is the deciding evidence. The writing pass took both files to zero mechanical hits across 160 prose lines: six passives given their actor, five -ly adverbs replaced, two of them by the number they were standing in for, and two headings rewritten that defined a thing by what it is not.
Four sentences carrying an anecdote about a boto3 bump and Postgres serialization, the eight-and-five counts, and a note that the detection costs nothing. The rule is the part a reader needs: a check name failing on several unrelated pull requests is the check being broken, and the output says which failures are shared. Two sentences. The counts it was quoting are in the recorded output further down the page anyway, where a reader can see the five marked lines rather than being told about them.
"which failures are shared" was passive in a sentence about who is to blame, which is the one place it should not be. The output marks those failures as the checks' own.
…ells "The cheapest call is the one nobody makes" was an epigram: a general truth shaped to sound quotable, carrying no fact a reader could use. The section around it had three more of the same class, which a word-level scan cannot see. A setup sentence, "Three things produce that funnel", existed only to introduce the list under it. Three bold headings in a row built a parallel triad. "Charge frontier prices for the privilege" was a sneer, and "the job only it can do" was a flourish standing in for "reading code". The section is now five plain paragraphs titled "What each stage costs", each one stating what happens and what it costs: ten removed by rules before any model runs, eleven judgements and two numbers for about 3,000 input tokens whoever answers them, a review on one pull request in eighteen, and the condition that the saving holds only while the cheap routes are right. Five more shape tells across both files, all of them defining a thing by what it is not: "the number to measure is not accuracy across the queue", "none of them says merge it now, and none of them merges anything", "the API base is a variable, not a glued-together path", "size is not review cost, and neither is line count", and "security_surface at 0.98 is not offset by infra_surface at 0.04". Each now states the positive version. Two passives the earlier pass missed, because the regex only looked for -ed participles: "the max of them is taken" and "that diagram is built from". Both files scan clean: 64 and 40 prose lines, no padding, no -ly adverbs, no passive in either -ed or irregular form, no corrective negation outside a table cell where "the author, who is not asking" is the plain meaning.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The pipeline diagram had grown to seven steps carrying every weight, threshold and rubric level. Accurate, but it read like a specification.
The flow diagram says the same shape in five boxes and one branch: a pull request, pre-triage on plain rules with no model, a fork where draft or a failing or running check stops and reports that state, otherwise one Jev call of eleven questions, the nine becoming an (effort, consequence) pair, and the pair picking one route. The route table now says what the repo owner does.
It ends with two worked examples so a reader can see what a skill running this reports:
Every figure comes from a committed report. The second example is from the gateway run because the airflow report no longer holds a
human-plus-author: the pull request that did is nowci-failing, so pre-triage removes it before Jev sees it. Both rows name their repository rather than mixing quietly.pipeline.htmlandpipeline.pngare deleted,render.pygoes back to one target, and the AGENTS.md layout line follows. No code change, so no release needed.