What Claude Code First shipped.
Every entry in docs/CHANGELOG.md, read at 54e1414 and dated by the commit that wrote it.
- CCF80 · 54e1414Rule pages print literal asterisks around every normative sentence that holds an italic term
Rule pages render a bold normative sentence holding an italic term as strong with em inside, and a test fails on any rule page that prints two asterisks.
- CCF79 · cc174ebThe site header links neither alegauss.github.io nor the GitHub repository, unlike every other alegauss site
Every page's header carries the "part of alegauss" chip to alegauss.github.io and a GitHub button to the repository, and a test fails on a header without either.
- CCF78 · 3e0abeemarkdownlint lints the governed docs/DEFERRED.md, and fails on a blank line roadkeep wrote there
markdownlint now ignores docs/DEFERRED.md beside the other four governed files, and its gate passes with the file as roadkeep wrote it.
- CCF77 · 72e902bThe guide and specification pages carry their own 892-line stylesheet, outside the design system
The guide and spec pages take fonts, tokens and site.css at 2026.3.22, re-keyed to the mark's rose, with analytics behind consent; style.css keeps only this site's own blocks.
- CCF76 · 84af454The site shows no mark of its own: no favicon, no logo in the header, no social card
Every built page links icon.svg and carries og:image, the guide header shows the mark, and a stdlib-drawn og.png renders the mark and title.
- CCF75 · 52a4b67the files-over-APIs open question names what would settle it, and nothing has been measured against it
F410 grades Shio's SH1141 at R1/S3: files cost 1.29x less and the agent writes 3.1x less; AP's open question and the guide say so, and corpus.md now lists a commit cited past a pin.
- CCF74 · cf33932the method admits no evidence of what a project's own sessions cost, so the session-budget question cannot be settled
method.md admits session-cost aggregates on four conditions, harness.md records H16-H20 and O4, and evidence/sessions holds a cost per task for all five projects.
- CCF73 · 14f3078the guide's lessons say nothing about what an agent-facing surface costs, and the page types its lesson count
The guide's ninth lesson shapes the agent's path to the task and measures it, linking F409, F8, AP-5 and AP-6, and its lesson count is computed from the cards.
- CCF72 · 04ef613no rule asks a product to benchmark canonical tasks through its agent surface against the generic path
AP-6 asks a product to benchmark canonical tasks through its surface and the generic path, gated on calls, tokens and the ratio, at SHOULD though F409 admits MUST.
- CCF71 · 6241956the files-over-APIs and the per-project session token budget claims are recorded nowhere as open questions
AP's open questions now hold files over an API for authoring, argued by Shio and never measured, and IS's hold a token budget on a project's own sessions, which no corpus project counts.
- CCF70 · 43b1502no rule says when a capability belongs on a CLI verb rather than on an MCP tool list
AP-5 lets a product offer a capability as a CLI verb where its agents have a shell, keeping MCP for shell-less clients and schema-checked input, at MAY because H9 disputes the cost.
- CCF69 · 289bdd2no finding records putting capabilities on a CLI rather than an MCP tool list where the consuming agent has a shell
F8 records Shio keeping three capabilities off its MCP tool list for a CLI its agent could reach, graded R2/S2, with winwright contrary where a schema guards a structured input.
- CCF68 · bdfd923no finding records that a task-shaped agent path cost a fraction of the generic API's tokens, or where the gap came from
F409 records that a task-shaped path cost Shio and freewilly a fraction of the generic API's tokens, graded R3/S4, and traces the gap to fewer reads rather than to REST.
- CCF67 · b8d1948the site opens on a table of rules, and a reader setting up a project has no plain-language way in
The site root is a plain-language guide with lessons linked to their findings and rules, and three charts built from the repository; the first site moved whole under spec/.
- CCF66 · b37c98dCI still excuses winwright from the pointer check after its pin was pushed
CI checks winwright's pointers against its remote, now that its pin is pushed; only the private Shio stays excused.
- CCF65 · d18bd46this repository declares no line terminators in .gitattributes, so it fails its own rule EP-2
This repository declares LF line endings in .gitattributes, claims level 2 in its own ccf.toml, and CI runs the conformance checker against it on every push.
- CCF64 · ebd1bbethe citation resolver takes about five minutes over 795 pointers, so it will be the gate a commit skips first
The resolver reads each corpus through one git cat-file batch process and the pin's ancestry set: 1,972 pointers take 1.5 s, down from 273 s.
- CCF63 · 3ccd1a7the resolver cannot check a quote that holds inline code or runs across comment-prefixed source lines
A quote is now checked as written: it may wrap onto the line before its pointer, keep backticks and run across comment lines, and a wrapped quote no longer passes unread.
- CCF62 · cbda3b6the repository has no README telling a newcomer what the spec is, how to read it and how to adopt it
README.md says what the specification is, where adopters, auditors and researchers start, the layout, the 0.x status and how to cite, linking rather than typing any figure.
- CCF61 · b138ae8the work has no license, no citation metadata and no statement of how it was authored
The text is under CC BY 4.0 and the code under MIT, each in its own license file, with the README naming which folders each covers.
- CCF60 · 07c9436an agent in another repository cannot consume the spec cheaply: no llms.txt and no page per rule
scripts/build_site.py publishes a page and a Markdown twin per rule and finding, llms.txt and the registry as data, built with no broken link in CI and deployed on demand.
- CCF59 · a868d89no process admits new evidence or a new rule, so the spec would either freeze or grow by opinion
spec/changing.md sets the path for a change: a proposal with rule, findings, level, cost and class, evidence through the protocol, a task and decision, deprecation before withdrawal.
- CCF58 · 7defbbethe specification has no version, so an audit cannot say which text it judged a project against
The specification follows semantic versioning of its obligations, with the version in one place and a tool that writes release notes from the rule diff and refuses too small a bump.
- CCF55 · cc9cd58shio has never been audited against the specification, so its conformance is asserted, not measured
Shio is audited at its head: 30 of 56 rules fail after the verifier rejected two scanner findings, achieved level 0, with proposed classes, offered tasks and cost recorded.
- CCF54 · cd31a2bwinwright has never been audited against the specification, so its conformance is asserted, not measured
winwright is audited at its head: 21 of 56 rules fail after the verifier rejected two scanner findings and added one, achieved level 0, with proposed classes and cost recorded.
- CCF53 · b75953efreewilly has never been audited against the specification, so its conformance is asserted, not measured
freewilly is audited at its head: 26 of 56 rules fail, one scanner finding rejected by the verifier and one added, achieved level 0, with proposed classes and cost recorded.
- CCF52 · d44c4d6polyweave has never been audited against the specification, so its conformance is asserted, not measured
polyweave is audited at its head: 19 of 56 rules fail with none confirmed false and 6 could not run, achieved level 0, with proposed classes, offered tasks and cost recorded.
- CCF51 · 19960eeroadkeep has never been audited against the specification, so its conformance is asserted, not measured
roadkeep is audited at its head: 22 of 56 rules fail with none confirmed false, achieved level 0, with proposed classes, offered tasks, circularity and cost recorded.
- CCF50 · 258e11ca project already Claude Code first that drifted has no path back except rereading the whole spec
adoption/realignment.md takes a drifted project from audit to classified failures, owner-approved tasks one level at a time and a diffed re-audit, and skills/realign performs it.
- CCF49 · 89ad3adan existing codebase adopting the practice has no procedure, and Shio's transition lives only in its history
adoption/brownfield.md takes an existing codebase from inventory to a declared level, verified against Shio's pin, with waivers for what history cannot meet, and skills/adopt performs it.
- CCF48 · dab72a5starting a project Claude Code first has no ordered procedure, so the order the five learned in is repeated
adoption/greenfield.md orders the install so each piece protects the next, from day-one decisions to the first audit, and skills/bootstrap performs it and stops before the first commit.
- CCF47 · 4912c35a new project has no starting kit, so each of the five wrote its own settings, hooks and skills
templates/ holds a conforming file per artefact citing its rules, and a repository assembled from them passes the checker at level 2 in a test (design superseded: hooks referenced, not copied).
- CCF46 · 65cfefda deliberate deviation from a rule cannot be recorded, so it reads as drift in every audit
An adopter's ccf.toml declares its level, spec version and waivers with reason, owner, date and end; the checker and audit report them as waived, or failed once expired.
- CCF45 · 54382acan audit report has no fixed format, so two audits of one project cannot be compared over time
An audit report is JSON in a fixed schema with a Markdown rendering, a verdict per rule where could-not-run blocks the level, and tools to validate, render and diff two reports.
- CCF44 · af9505cthe rules that need judgement cannot be audited without an agent reading the whole spec
The repository ships an audit plugin: a skill that runs the checker then a scanner per chapter and a verifier for judgement rules, validated strictly in CI with a pinned CLI.
- CCF43 · 1251babautomatable rules have no checker, so each audit rederives by reading what a script could measure
scripts/check_conformance.py decides six rules from a repository's files with a third verdict, exits non-zero below the claimed level, and passes and fails its fixture repositories.
- CCF42 · 4c3bbd2traceability is unchecked: a rule citing no finding, or a finding citing no evidence, would pass
A CI gate traces every rule to findings whose grade admits its keyword and every finding to a primary pointer, and regenerates spec/traceability.md, the matrix.
- CCF41 · e6ef31arules exist only as prose, so no tool can list them, filter them by level or check a repository against them
spec/rules.toml lists all 56 rules with keyword, level, statement, findings and harness facts, generated from the chapters with a CI check, and a query command answers without reading them.
- CCF40 · 4978608conformance has no levels, so a project cannot say how far it adopts the spec or what to do next
spec/conformance.md defines three cumulative levels, governed, gated and measured, plus the agent-facing profile, assigns all 56 rules with reasons, and says how a claim is made.
- CCF39 · f2a41edthe practices that resolved those failures are described only inside the projects that invented them
spec/catalogue/patterns/ describes 14 practices that resolved the failures, each with the anti-pattern it answers, its evidence, its rules and the cost it adds.
- CCF38 · 6d5e935the recurring failures have no names, so a reviewer cannot say which one a project is committing
spec/catalogue/anti-patterns/ names 15 recurring failures, each graded by its best finding and linked to the rules that prevent it and the pattern that resolved it.
- CCF37 · 3f7e4f7no form fixes how a pattern or anti-pattern is written, so entries would mix problem, remedy and evidence
spec/catalogue/README.md fixes the pattern and anti-pattern form, and a CI check holds every entry to its sections, bars RFC 2119 keywords and requires a rule link.
- CCF36 · 7e77decno normative chapter covers the prose an agent writes, though model-written prose was measurably detectable
spec/AW.md states three rules on agent prose: generated or checked figures (MUST, F7), length limits refused at the write (MUST), and gated style rules (SHOULD, per D7).
- CCF35 · 4fd867ano normative chapter covers several agent sessions working in one repository at once
spec/CS.md states five rules for sessions sharing a checkout: claims on task and paths, staging within the claim, disturbed gate runs as could-not-run, no cited figures from a busy tree.
- CCF34 · 0323717no normative chapter says what the person decides and what an agent must never certify for itself
spec/HR.md states that an agent does not close work needing a person's judgement or absent hardware, and applies a stored human bar instead of its own margin.
- CCF33 · ecc0b37no normative chapter covers projects that ship tools, skills or plugins for other agents to use
spec/AP.md states four rules for products agents use: gated token ceilings, names checked against the catalogue, tests from the published artefact, adoption proved in a consumer.
- CCF32 · dc5753fno normative chapter covers where a session runs: web versus local, shells, encodings and line endings
spec/EP.md states six portability rules bounded to Windows: a declared tool copy, pinned line endings, named encodings, BOM handling, bytes before blame and no heredoc edits.
- CCF31 · fcbcedcno normative chapter states a permission and safety posture, and the corpus runs with prompts bypassed
spec/PS.md states three permission rules the evidence carries: name what replaces removed prompts, refuse forbidden actions in configuration, keep credentials out of the tree.
- CCF30 · 6da086cno normative chapter says which rules must be enforced by hooks rather than by instructions
spec/GH.md states six rules on what the harness enforces: broken rules leave prose, a CI gate behind every guard, every route matched, silent failures surfaced, configuration gated.
- CCF29 · 3c3ce10no normative chapter says what an agent must run, keep and report before calling work done
spec/VG.md states nine rules on what must run and pass before work is done, from gates on the committed change to the third verdict, CI on every push and expiring red exceptions.
- CCF28 · 4c9b2c1no normative chapter says how an agent turns finished work into commits
spec/CD.md states six rules on turning work into commits: one task per commit, a gate between tasks, staging by path, the agent's own title, no empty filings, and a stated attribution policy.
- CCF27 · bea3352no normative chapter says how an agent plans, records and closes work
spec/PG.md states six planning rules satisfiable by any tool: governed files, symptom-first tasks, recorded false premises, rationale moved before deletion, criteria-closed blocks and deferral.
- CCF26 · 0392668no normative chapter says what may load on every turn and what must be trigger-loaded instead
spec/IS.md states six rules on the instruction surface, each capped by its findings: a gated budget on any every-turn file, bounded skills and read-on-demand files, checked name lists.
- CCF25 · f08c43aterms such as every-turn file, governed file, gate, guard, ledger and verdict are used without definitions
spec/glossary.md defines 18 terms, settling the corpus's two senses of a block, and a CI check fails any italic term in a rule sentence that has no entry.
- CCF24 · 64d3588the specification has no frame: no chapter order, no rule addresses and no conformance keywords
spec/ has its frame: scope, audience and chapter list in README.md, and in conventions.md the RFC 2119 keywords tied to evidence grades, the rule form and permanent addresses.
- CCF23 · e4d9342no cross-project metrics exist, so claims about scale, cadence and context cost are adjectives
scripts/corpus_metrics.py computes commits, cadence, commit conventions, ledgers, every-turn bytes and tests for all five pins into committed data, methods stated.
- CCF22 · 2b3b735the drift already present in the corpus is uncatalogued, though each case shows prose rules decaying
evidence/drift.md inventories nine verified contradictions inside corpus projects at their pins, with dates, days standing and whether a check could catch each.
- CCF21 · 4af3519the projects disagree on attribution, instruction files, permissions and CI scope, and nothing settles it
evidence/divergences.md settles seven divergences against graded findings, allowing a documented deviation where evidence cannot decide (design superseded: one register).
- CCF20 · 38382d8lessons are scattered over five ledgers, so one is counted five times and a contradiction not at all
evidence/findings/ holds 37 graded findings across five topics, each merging observations by independent origin with resolving pointers, and a generated index CI checks.
- CCF19 · 22e8c25no case study describes shio, the only project that adopted the practice after years of history
evidence/cases/shio.md tells Shio's brownfield case in the shared structure with 114 resolving pointers, splitting before and after adoption by commit rather than by date.
- CCF18 · 177e401no case study describes winwright, whose agent surface is shipped for other repositories to use
evidence/cases/winwright.md tells winwright's case in the shared structure with 107 resolving pointers, and corrects the design on WW69, the reopened blocks and the red pushes.
- CCF17 · ec0bf8cno case study describes freewilly, the project that works without any instruction file
evidence/cases/freewilly.md tells freewilly's case in the shared structure with 103 resolving pointers, and corrects the design on CONTRIBUTING, DD115 and the web-session commits.
- CCF16 · fa2c498no case study describes polyweave, the one project governed from its very first commit
evidence/cases/polyweave.md tells polyweave's case in the shared structure with 112 resolving pointers, measuring throughput from the ledger and correcting the design on 871a15b.
- CCF15 · 3ae5d64no case study describes roadkeep: its timeline, agent surfaces, gates and the incidents that shaped them
evidence/cases/roadkeep.md tells roadkeep's case in the shared nine sections with 121 resolving pointers, correcting the design where the pinned sources disagree.
- CCF14 · 4ef901fno related work is surveyed, so the spec cannot say what is new and what restates ADRs, RFC 2119 or pattern languages
evidence/related-work.md surveys ADRs, RFC 2119, patterns, Diataxis, commits, method, Anthropic's guide and four agent studies, and names five claims with no precedent.
- CCF13 · 2311c25the Claude Code behaviour the spec relies on is cited from memory rather than from versioned documentation
evidence/harness.md registers 14 documented harness facts with page, date and CLI 2.1.280, and 3 observed ones with evidence, one now contradicted by the docs.
- CCF12 · f7abda0threats to validity are unstated: one author, one harness, one model family, five projects over six months
evidence/validity.md states the construct, internal, external and reliability threats, traces the direction of copying from dated commits, and reports the measured error rate.
- CCF11 · 48621afthe field notes were extracted by agents and not one of their claims has been checked against the pinned sources
All 564 field-note claims are marked against the pins, 789 pointers resolve, and the error rate is recorded: 13.8% overall, 12.9% to 16.3% per project.
- CCF10 · 9bc921dclaims carry no strength, so one incident in one project reads like a pattern seen in all five
evidence/grading.md grades findings on recurrence and strength, counts copied practice once, maps grades to MUST, SHOULD and MAY, and sets itself against EBSE levels.
- CCF9 · 632954ean evidence pointer has no grammar, so nothing can check that the path, line or commit it names exists
A pointer grammar is defined, and a tested resolver checks each pointer's project, commit, path, lines and quote against the pinned corpus in CI (design recorded in
evidence/method.md). - CCF8 · 6d9afc8the corpus is named by path only, so a citation breaks as soon as a source repository moves on
evidence/corpus.md pins each project to the full commit the notes read, each confirmed as HEAD during extraction, with remotes, work outside the pins and descriptive data.
- CCF7 · 2fe83e6no protocol says how evidence is gathered or judged, so a finding here cannot be reproduced or challenged
evidence/method.md states the research questions, case design, sources, extraction queries, coding scheme and disagreement rule; the first five prompts are kept verbatim.
- CCF6 · 80ba574a Write over an existing file is accepted silently here, the path that destroyed about 700 lines in pportal
A Write over an existing non-empty file is now refused by a PreToolUse hook naming Edit, vendored from polyweave with its source commit in the header.
- CCF5 · ffcf35bthe every-turn files carry no size budget, so they can grow the way Shio's agents.md reached 186 KB
agents.md and .claude/CLAUDE.md now carry line and byte budgets lint enforces, and every skill written here is capped at 700 characters of description and 6,000 of body.
- CCF4 · 6ba7fffnothing states how work is committed here, so batching, stray files and mislabelled docs commits are all possible
Committing here is a trigger-loaded ccf-dev skill: one task per commit, staged by path, message from a file, the three gates first (design superseded: staged by path).
- CCF3 · 23583e7no gate runs on push, so a lint failure or a broken link in the specification reaches main unremarked
Every push and pull request now runs roadkeep lint, an offline link check and a Markdown rule set; both new gates were seen failing locally (design superseded: lint is in roadkeep.yml).
- CCF2 · a247bd0the governed docs have no guard here, so a hand edit of ROADMAP.md or IMPROVEMENTS.md passes unnoticed
A hand edit of a governed doc is now denied by a committed PreToolUse guard naming the roadkeep command to use, on every clone, with a merge driver for the four files.
- CCF1 · 46578e7a session opening this repository is told nothing: no every-turn file states purpose, language or write paths
A session opening this repository now loads agents.md through .claude/CLAUDE.md, stating the purpose, three laws on evidence, English and roadkeep writes, and the layout.
- CCF57the claim that conformance reduces cost has no before-and-after measurement on any realigned project
roadkeep's owner accepted the drift tasks, and the 14 days before realignment are measured by a script that measures the window after the same way.
- CCF56nothing shows the audit is accurate: its false findings and misses are unmeasured
validation/accuracy.md reports that the verifier confirmed 107 of 112 scanner findings, says why that is agreement not accuracy, and holds the sample for the person.