All notable changes to this repo. Format after Keep a Changelog 2.0.0; versions follow SemVer. Dates are the day the change was pushed.
Every entry names the adversarial audit round that produced it where one did. Audits are read-only passes against the stated release scope; every finding is reproduced before its fix.
Unreleased
0.10.1 - 2026-11-18
Fixed
- Respect
CODEX_HOMEfor--global --ai agents(install, re-run, atomic rollback,learnand--uninstall), pre-selection and session discovery, and warn without modifyingAGENTS.override.mdwhen a non-empty override shadowsAGENTS.md(#51). - Replace blanket ChatGPT file-access copy with surface-specific wording, and describe Codex memory notes as local to the machine or connected host (
~/.codexor$CODEX_HOME) rather than global to an account (#52). - Document the enclosing Git
info/exclude(and level-3--precommit) exception to the project write boundary inREADME.md(#53). - Align
docs/learning.mdanddocs/paste-guide.mdwith full declined-entry ChatGPT exports andOVER BUDGET/--strictbehavior, and scopedocs/companions.mdenforcement claims to operations routed through the configured companion (#54).
0.10.0 - 2026-11-18
Added
- Add
templates/GEMINI.mdpointer headers,## binding:geminirule sections,.gemini/commands/personalize.tomlfor/personalize,hooks/gemini-cli/session-start.shemittingSessionStartadditionalContextJSON, and Google Antigravity (~/.gemini/config/GEMINI.md, 24,000-byte cap) documentation (#48). - Read Codex
memories_v2/(MEMORY.md,memory_summary.md,rollout_summaries/*.md) alongside legacy~/.codex/memories/(#46). - Ship
skills/agent-personalizer/SKILL.mdin the npm package (#50).
Changed
- Require Node.js 22 or later and run CI on Node 22, 24 and 26 (Active LTS) (#49).
- Pack
examples/answers.jsoninstead of full rendered example subdirectories, remove unused pre-0.8.0 interview exports fromrender/onboarding.cjs, share CLI helpers inbin/cli-util.cjs, and maketest/rollback.test.jsself-seeding (#50). - Document ChatGPT's 5,000-character paid and 1,500-character Free/Go limits, update
--globaland Hermes Agent target lists across docs and issue templates, add/docs/guides/section anchors, and expand CLI--helpsynopses (#49).
Fixed
- Audit 2026-11-17: resolve Gemini CLI
v0.42+project slugs fromprojects.jsonand.project_root(including subfolders and~/.cache/.gemini/tmp), and reconstruct Gemini CLIv0.40+append-only.jsonl(MessageRecord, string and part-arraycontent,$patch,$rewindTo),session-*.jsonfiles andsubagentfiltering (#46). - Audit 2026-11-17: normalize Unicode to
NFCincheck/gate.cjsand collapse whitespace inbin/learn.cjspath redaction (#47). - Audit 2026-11-17: scope
bin/global.cjspath qualification to installer boilerplate and rules so userUSER.mdprose andLEARNED.mdentries are never rewritten, and isolateHOMEinscripts/record-demo.sh(#47). - Audit 2026-11-17: sync and pin
website/release-fallback.jsontopackage.json, fix thedocs/install.md--globaltarget list, and tailor thelearn --globalcompletion hint (#48, #49).
0.9.1 - 2026-11-10
Security
- Harden release automation, HTML sanitization and privacy checks.
0.9.0 - 2026-11-04
Added
- Add opt-in compact rule rendering and a warn-only instruction file size budget.
- Add an opt-in level-3 git pre-commit hook for the forbidden-string gate.
0.8.7 - 2026-11-03
Changed
- Name Hermes Agent among the supported AIs in the package description, README, llms.txt, the skill description and the project page.
0.8.6 - 2026-11-03
Added
- Add Hermes Agent as a global target with a marked block in
~/.hermes/SOUL.md; project installs use theagentstarget.
0.8.5 - 2026-10-29
Fixed
- Audit 2026-10-29: find Claude Code sessions for project paths with spaces, dots and underscores while retaining legacy folder names.
- Audit 2026-10-29: skip whole messages and Codex memory notes that match conservative secret patterns, reporting only the skip count.
- Audit 2026-10-29: apply the 2,000-character message cap to Codex memory notes, retaining private-term filtering and the 20-file cap.
- Audit 2026-10-29: disclose global Codex memory scope and secret skips in consent, and document both project and global private lists.
- Audit 2026-10-29: preserve intra-word underscores in read-first filenames while escaping hostile Markdown.
- Audit 2026-10-29: refuse local notes paths containing .git, .github or node_modules segments, ignoring case.
- Audit 2026-10-29: remove the uninstalled folder-template path from rendered home files and refresh all three examples.
- Audit 2026-10-29: register the project session-start hook through CLAUDE_PROJECT_DIR so it runs from subfolders, and explain user-level absolute paths.
- Audit 2026-10-29: align memory status, global learning help, level errors, AI target contribution steps and symlink guidance with supported behaviour; remove the inaccessible audit-brief citation.
- Audit 2026-10-29: warn that profile and off-limits names enter home files that may be committed.
- Audit 2026-10-29: use short temporary bases for session-path fixtures so deep TMPDIR paths do not cause false failures.
0.8.4 - 2026-10-28
Added
--globalapplies a level-1 profile across Claude Code, Codex and Gemini CLI projects, with byte-preserving marked blocks, safe uninstall and shared atomic rollback (#29).
Changed
- Levels 1 and 2 render the Core 5 rules; level 3 keeps all eleven and editable rule sources. Untouched older generated files refresh through recorded ownership; edited files stay for review (#33).
Fixed
- The 0.8.4 follow-up audit adds the Windows
USERPROFILEfallback and home-free project AI defaults, preserves outside edits when uninstalling adopted pre-existing targets, and checks Markdown links after blank lines and escaped backticks. - The 0.8.4 follow-up audit fixes bound Claude and Codex session reads to contenders for the newest 20, validate and clamp message timestamps to file mtime, refuse foreign global blocks and invalid HOME values before writes, adopt explicitly re-rendered Core 5 upgrades, and keep code examples containing HTML comment openers from hiding later Markdown links.
- Session learning selects the newest 20 usable sessions by latest message timestamp across all sources, including legacy Gemini logs, with file mtime used only when a session has no valid timestamp (#38).
- Digest removal documentation now agrees with recorded-hash ownership and preserves edited or unrecorded files (#39).
- Markdown link checking strips nested and unterminated HTML comments completely (CodeQL
js/incomplete-multi-character-sanitization). - Security reporting links directly to private advisories and states the existing acknowledgement and fix windows (Scorecard Security-Policy).
- Publishing uses Node 24 and checks bundled npm is at least 11.5.1, replacing the mutable global npm install while retaining the pinned setup action (Scorecard Pinned-Dependencies).
0.8.3 - 2026-10-19
Changed
- The README and llms.txt "Part of a set" section now renders from the shared set file. The README table has four rows, adds repo-build-skill, and uses each package's npm description. CI checks the rendered blocks and locked file hashes for drift offline.
0.8.2 - 2026-10-18
Changed
- The model-orchestrator row in "Part of a set" and llms.txt matches its current npm description.
0.8.1 - 2026-10-17
Fixed
- Re-render keeps user text outside generated blocks out of whole-file ownership updates and leaves legacy configurations without
installedon the legacy removal path, so uninstall retains added instructions and removes untouched legacy targets (Codex audit 2026-10-17, F1; agy audit, A3). - Local notes paths inside tool-owned paths are refused,
learnpreserves edited or pre-existing.agent-personalizer/.gitignoreanddigest.mdfiles without adopting their hashes, and uninstall keeps unrecorded or edited digest-folder files (Codex audit 2026-10-17, F2; agy audit, A5). - Renderer ownership configuration writes share target staging and rollback, with clean diagnostics on failure (Codex audit 2026-10-17, F3).
- Session and Codex memory privacy filtering strips control characters before matching and checks the final text after context removal, including private terms rejoined across newlines, item boundaries or spaces (Codex audit 2026-10-17, F4; agy audit, A1 and A6).
- ChatGPT exports every declined entry in full and reports overflow;
--strictrefuses over-budget lists before writing (Codex audit 2026-10-17, F5). - Render, drift checks and uninstall strip a leading UTF-8 BOM before fence, marker and
LEARNED.mdparsing while preserving it byte for byte on output (Codex audit 2026-10-17, F6; agy audit, A2 and A4). - The demo recorder follows the current folder, AI selection and default-No session-consent prompts (Codex audit 2026-10-17, F7).
- Uninstall restores a pre-existing home file byte for byte: it removes the blank separator line install added along with the block. An untouched
LEARNED.mdtemplate is removed, and.agent-personalizer.jsonis removed when only stripped home files remain (#32). - Inside a git repo, install adds
USER.md,LEARNED.md,.agent-personalizer.jsonand the ChatGPT paste files to.git/info/excludein one marked block per install folder, and notes that the home files carry your profile. Uninstall removes the block (#30). - Learning reads sessions from the current project by default;
--all-projectsexplicitly includes sessions from other projects (#31). - The F5 historical-install check skips when its version tags are absent, including in shallow clones, and reports the skip separately from a pass (#34).
0.8.0 - 2026-10-06
Added
- Consent-gated local readers for recent Claude Code, Codex and Gemini CLI sessions, plus Codex memory notes. Readers stream bounded input, keep typed messages, skip automation and unsafe paths, and filter private-list matches without printing message text.
learncollects new evidence after a yes;learn --forgetdeletes the local digest without touching AI app history. The digest labels messages as data and stays in a gitignored local folder.LEARNED.mdat every install level owns confirmed, proposed and declined entries. Confirmed entries reach every personal AI target, declined entries tell the AI never to ask again, and--checkdetects unrendered entry edits.- One shared procedure asks up to five questions with counts and quotes from an unhandled digest. Claude Code receives
/personalize; Codex and Gemini CLI receive the same flow in their home files. - Four rules with origin stories: Ask before it repeats, Their words point at the thing, Symptoms are not causes, and Surface what they can't see.
- Learning and privacy documentation, synthetic session fixtures, and two more invented example users with different entries.
Changed
- The first interactive install asks only for the folder when needed, selected AIs and consent. AIs found locally are pre-selected, level 1 is the default, and session consent defaults to No. The personal questionnaire is gone;
--fulland--quickremain accepted no-ops. - Re-runs ask nothing and preserve saved settings, including legacy name, pronouns, work and focus answers. Non-default legacy identity values remain in the generated profile's "About me" section.
- Uninstall always keeps
LEARNED.mdand removes the digest. Non-interactive installs never read sessions; scripted consent useslearn --yes. - Privacy guidance explains local session access, AI provider visibility, other people's details, work accounts, private-list filtering and undo.
0.7.2 - 2026-10-06
Changed
- The Contributor License Agreement defines a Contribution as work submitted through a pull request, matching the pull request checkbox that records agreement.
- The README Contributing section links the Contributor License Agreement and explains how agreement is recorded.
0.7.1 - 2026-10-06
Changed
- The README opens with a 14-second promo trailer in the same format as the sibling repos. The terminal recording of the interview moves to the install guide.
- Both GIFs load from GitHub, and GIFs no longer ship in the npm package.
0.7.0 - 2026-10-06
Fixed
- An interactive re-run shows your saved answers as the defaults and keeps saved values for questions it does not ask, so Enter no longer resets answers or clears the off-limits list (found 2026-10-05).
- ChatGPT's profile export now follows edited and pre-existing USER.md files. Profiles byte-identical to the default generated profile keep their compact bytes; changed generated identity fields compact conservatively, and custom text stays in full with the existing budget warning (Codex audit 2026-10-04, F1).
- Re-runs migrate untouched home-file pointer lines to the new notes path. Edited lines stay byte for byte and are named for manual review; the previous notes folder stays in place and in the uninstall inventory (Codex audit 2026-10-04, F2).
- Changed answers refresh untouched notes templates and rule copies. Turning signing off removes unchanged signing pointers, template lines and the copied rule; edited copies stay and are named (Codex audit 2026-10-04, F3).
- Uninstall checks every path recorded by any install run, including earlier notes paths and conditional rules. Recorded hashes decide removal, edited files and user notes stay, only recorded install-created empty directories are removed, and the configuration goes last (Codex audit 2026-10-04, F4).
- Newer installers refresh untouched runtime tools and rules. Recorded hashes or a shipped table of earlier release hashes recognize untouched copies; edited tools stay and the log gives their known origin and replacement step. The configuration records tooling versions (Codex audit 2026-10-04, F5).
- Installer writes and removals are planned together before commit. A failed install restores changed files, permissions and new empty directories (Codex audit 2026-10-04, F2 to F5).
- Relative links in shipped Markdown resolve inside the npm package. Root contributor, security, maintenance, release and support pages now ship, and the install guide links a shipped answers example (Codex audit 2026-10-04, F9).
- A successful renderer
--checkis silent, matching the documented bundled-example verification command. Drift and refusal diagnostics retain their exit codes (Codex audit 2026-10-04, F10).
Added
- Harness checks 92 to 96 reproduce the five installer findings, including edited profiles, pointers, notes and tools. Check 97 verifies packed Markdown links (Codex audit 2026-10-04, F1 to F5 and F9).
- Uninstall unit regressions cover historical inventory, symlink refusal, unrecorded files and directories, invalid ownership records and installer rollback. Package-copy fixtures exclude the private forbidden list (Codex audit 2026-10-04, F4 regression work).
- CI tests Node 24 alongside 18, 20 and 22. Its stable aggregate job,
Repository checks, requires every other workflow job to succeed; Windows also exercises re-runs and uninstall (Codex audit 2026-10-04, F7 and F8). - npm funding links to GitHub Sponsors. MAINTAINERS.md, SUPPORT.md and a question form provide the owner, release and support routes (Codex audit 2026-10-04, F9).
Changed
- The first screen uses the selected profile-and-rules wording, a parts table and activation steps. Common questions precede the docs index; update and removal summaries describe ownership and history, and llms.txt follows the same lead (Codex audit 2026-10-04, F6 and F10).
- Codex guidance describes root-to-working-directory merging, override files and the default guidance cap, with a current official source. ChatGPT activation uses the current interface's fields and displayed limits. Installation recommends Node 22 or 24 while retaining 18-or-later compatibility (Codex audit 2026-10-04, F7).
- The npm description names the interview, profile, rules and supported targets. The custom social-preview image is removed in favor of GitHub's default link preview (Codex audit 2026-10-04, F9).
0.6.5 - 2026-10-04
Changed
- LICENSE names Auny LLC as copyright holder; outside contributions now require the CLA in CLA.md.
0.6.4 - 2026-09-30
Changed
- CLAUDE.md now imports AGENTS.md so Claude Code loads the same contributor rules as every other agent from one file.
- The feature form asks for an observable success check.
- AGENTS.md names the re-render command.
0.6.3 - 2026-09-24
Added
- CI installs the packed tarball in a fresh consumer project on Linux and macOS, checking levels 1 and 2, ChatGPT paste files and the installed version (#26).
0.6.2 - 2026-09-24
Changed
- npm keywords now match the GitHub topics, so npm search finds the package by the same terms:
claude-md,gemini-md,system-prompt,ai-memory,llm,cli,mcp,obsidianandprompt-engineeringadded (11 to 20). No code or output changes.
0.6.1 - 2026-09-23
Changed
- README install note drops the GitHub-source command; the npm command is the install path.
docs/install.mdshows the standard way to run one release:npx agent-personalizer@<version>.
0.6.0 - 2026-09-23
Presentation and safe removal from the repository presentation plan.
Added
--uninstall --dir <folder>compares generated files with the current render of stored answers and sources. Edited files are kept and named. Pre-existing home files retain their own bytes while an unchanged generated block is removed.--drypreviews the removal plan. Uninstall preflights tracked paths, refuses symlinks and traversal, removes only unchanged scaffold and tool copies, and removes the configuration last while keeping it for retained tracked files. The hook registration cleanup is printed as a manual step.- Original home ownership is recorded in
.agent-personalizer.jsonaspreservedTargets. Content is checked by rendering and byte comparison. Older configurations retain files whose ownership cannot be established from an exact generated-file match. - An uninstall regression runner covers clean levels, edited files, pre-existing homes, reinstallation, local rule changes, user notes, path refusals, dry runs and missing configuration. The shell harness now reports 91 checks.
llms.txt, a documentation index, installation and guarantees guides, and a published-package demo recording script.
Changed
- README leads with the interview and generated files, shows real example excerpts, documents AI loading paths, links both sibling tools and explains removal. Detailed flags and instruction tiers have their own guides.
- npm description names the generated profile and rule files.
llms.txtjoins the published files list. - Contribution and maintenance guidance names the actions readers can take.
Fixed
- Repository
npm run renderandnpm run checktarget the maintained example. Contributor instruction files at the repository root remain hand-written. - Installation copy accurately describes file selection, conditional interview questions, local and cloud notes, signature opt-in and the mechanisms level.
- Install now records a SHA-256 of the exact bytes it writes for each tracked file, and of the marker block alone for a pre-existing home file, in
.agent-personalizer.jsonunderinstalled.--uninstallcompares against that record first, so a later release of this very package (an unpinnednpx agent-personalizer --uninstallafter an upgrade) can no longer make an untouched file read as edited and get kept forever; a configuration from before this field existed still falls back to the previous comparison (code-reviewer audit,AUDIT_BRIEF.md). - The former routing placeholder now links to model-orchestrator.
--level 4continues to install the same files as level 3 for existing scripts.
0.5.2 - 2026-09-17
One Codex read-only round covered the export together with its first consumer. It returned no finding against the export: check 90 passed and the shared DIE_THROWS state produced no failure in the installer's call path.
Added
- The rule parser is a library.
parseFrontmatterandparseSectionsare exported fromrender/render.cjs, so another renderer can read the same rule-file format with the same validation instead of re-implementing it. The first consumer renders one person's CLAUDE.md sections and session-start hook bodies from rule files (the maintainer's own stack). Check 90 asserts both exports parse a shipped rule and refuse a bad section as aRefusal. The harness is 90 checks now.
0.5.1 - 2026-09-12
Issue #25, filed on a recheck of 0.5.0 by the same outside reader as #19 to #24. Reproduced, then fixed, then handed to Codex as one adversarial read-only round against AUDIT_BRIEF.md, which returned six findings. All six were reproduced by hand and fixed here, and every fix has a harness assertion inside check 89 that goes red when the fix is backed out. The harness is 89 checks now.
Fixed
- The two files rendered from the onboarding answers no longer name a notes path that is not on disk (#25). #19 fixed the home files: a level-1
CLAUDE.md/AGENTS.mdcollapsed fournotes/...pointers into one line pointing atAGENT_ONBOARDING.md. The file that line points AT, andUSER.md, still namednotes/README.md,notes/sessions/,notes/decisions.mdandnotes/inbox/, so the first session was told to read and write into a folder the installer had not created. One predicate now owns the question for all of it,onboarding.notesPending(): a local notes folder is intended for every non-cloud tool, and the level-2 scaffold is what creates it, so below level 2 the renders say the folder is absent, tell the AI to propose writes and create nothing unasked, and defer the write policy to "once it exists" rather than dropping it. A cloud notes tool is untouched at any level, because it never gets a local folder. The session-start contract and both ChatGPT boxes carry the same pending line, compressed: it is never longer than the line it replaces, which the harness asserts, because the ChatGPT box budget has about two characters of slack on this repo's own fixture. - The disk is the truth about whether the notes folder is there; the level was only a proxy (round 1, findings 3 to 5). The first cut of the #25 fix keyed on the install level, which is wrong in both directions: a hand-written
.agent-personalizer.jsoncarries nolevelat all, so a plainrender.cjsrun emitted livenotes/README.md, session-log, decisions-log and inbox pointers into a folder that did not exist; a level-1 install landing in a folder that already had a notes folder told the AI that folder was absent; and a re-run at a lower level does not delete what a higher one created, so a 2-then-1 sequence rewroteUSER.mdto call a folder absent whileAGENT_ONBOARDING.mdstill named its paths. One boolean now answers the question for the home-file collapse and both rendered files, from the presence of the folder README, which is the file the AI is told to read before it writes. The renders stay pure functions of the answers plus that boolean, becauserender/onboarding.cjsis copied into level-3 installs and the look at the disk belongs in the caller. ask-before-every-writekeeps "show what you would write and where" in the pending contract line (round 1, finding 1). The compressed line had dropped it, and that clause is the consent: someone who pasted the two ChatGPT boxes and nothing else had no other copy of it. The three pending lines now each carry "create nothing unasked" plus their own policy, and all three are shorter than the line they replace; the harness installs every policy at level 1 and refuses an over-budget box, rather than asserting a proxy for it.- A read-only or unknown notes tool is no longer told its own notebook arrives at level 2 (round 1, finding 2). The pending section said "Notes will live in Microsoft OneNote ... once that folder exists", but level 2 creates the local fallback folder, not a OneNote notebook: only the fallback is pending, and the notebook exists already.
- The stale-notes notice no longer keys on one sentence (round 1, finding 6). Editing the sentence it matched silenced it while the other stale line stayed. A kept
USER.mdthat does not name the folder README while the folder is there now earns the notice however it was worded. - A
--level 2run puts the real paths back inUSER.mdtoo (#25).AGENT_ONBOARDING.mdis regenerated from the answers on every render, so it repointed itself once the renderer knew the level;USER.mdis yours once it exists, so it is rewritten only when it still equals the render of the same answers at the previous level, byte for byte. If you edited it, it is kept and the run says which two lines are now stale, in place of repointing it underneath you.
Changed
- Three harness cases moved from level 1 to level 2 (#25). Checks 60, 64 and the obsidian-tc contract assertion install at level 2, because every path they assert on has to exist on disk to be named at all; their subject is the
notes_tooltable and the ChatGPT budget, not the install level. Check 89 covers the level-1 shape, including a mechanical loud negative: at level 1 every line naming a path under the notes base must also say the folder is not there yet, so a new line cannot quietly reintroduce a dead pointer.
0.5.0 - 2026-09-08
A first-run walkthrough of 0.4.3 by an outside reader, filed as issues #19 to #24. Every one was reproduced before its fix, and each has a harness check that fails on 0.4.3 and passes here (checks 84 to 88).
Changed
- The interview is short by default (#23). A bare
npx agent-personalizerasks the seven questions that change behaviour, not all 23.--fullasks the rest.--quickstill works and is now the default it used to name. - A conditional question is asked only when it applies (#23). The Obsidian question is asked about an Obsidian vault and nothing else; the "name your tool" question only when the tool is
other. That answer was previously asked of everyone and skipped by--quick, so a short interview could pickotherand never name it. One exported rule,onboarding.asks(), decides this for the installer and for the harness, so the check tests the installer rather than a copy of it. - A level-1 install no longer names a notes folder it did not create (#19). Level 1 writes four files and no folders, and its
CLAUDE.md/AGENTS.mdsaid to readnotes/README.md,notes/sessions/,notes/decisions.mdandnotes/inbox/: four dead pointers in the first session. The home file now carries one line pointing atAGENT_ONBOARDING.md, and a later--level 2restores the four real pointers in the same run that creates the folder behind them, touching only lines this installer wrote. - One verb per file in the install log (#20). The installer wrote a home file from its template and the renderer then filled its marker block, and both steps printed
wrote; a re-run printedkeptand thenwrotethe same file. The renderer now says what is true of that file:wrotea file that did not exist,update ... (rendered block)when the block changed,ok ... (rendered block already current)when it did not. The installer labels the template writewrote CLAUDE.md (pointer file; the renderer fills its block below). - The success text names the folder that was installed, and the paste step (#21). The rerun command printed
--dir ., which writes into wherever it is pasted from, not where the install went; it now prints the absolute path. The same block separates the files an AI reads from the folder by itself (CLAUDE.md,AGENTS.md,GEMINI.md) from the ones a human still has to paste (claude.ai, ChatGPT, a system prompt), and linksdocs/paste-guide.md. The companions line prints only when more than one AI was installed, which is what it is about. - Level 4 is no longer offered (#22). It was listed as a peer of 1 to 3 and then described as reading material for a layer that ships nothing. The installer, its
--help, and the README now stop at 3.--level 4is still accepted so an older script keeps working, and says outright that it installs exactly what level 3 installs. - The README stops mixing levels with tiers (#22). The install table has three rows and one name for the choice; the five tiers are introduced as where instructions live once written, not as something to pick. The heading "Level 1: one file" said one file over a table that said four; it now says what it writes.
--yesand--defaultsare documented as the different flags they are (#23).--defaultsis the answer source,--yesis non-interactive mode and needs--dir,--aiand--level.
Fixed
- The default profile told the AI to "settle facts itself" under a heading addressed to it (#24). The interview asks the human about the AI, so its option labels are third person;
USER.md,AGENT_ONBOARDING.mdand the session-start contract address the AI, and now take a second-person phrasing where a choice carries one ("settle facts yourself"). Every default install shipped the wrong voice, and the same line was copied into the ChatGPT boxes. - Angle-bracket placeholders no longer vanish on GitHub (#24).
Last edited by: <ai> <model> <date>rendered asLast edited by: ·in the four note templates, because GitHub treats the brackets as HTML. They, and five more in the same files (# <Folder name>, the week heading, the decisions-log example row), are now inside code spans. The README's own two occurrences were already fenced and rendered correctly; the files it links to were not. Check 88 scans every template, skipping fenced blocks and HTML comments, so a new one cannot reappear.
Added
- Harness checks 84 to 88, one per issue, each verified to fail against the 0.4.3 tree and pass here. 88 checks in total.
0.4.3 - 2026-09-06
Added
authorinpackage.json, so npm shows a byline:aunysillyme (https://github.com/aunysillyme). It needs anallow:entry in the private forbidden list, which is the gate working as designed: attribution is opt-in, one exact string at a time.- An npm version badge in the README, matching the other published package.
- The installer's last line now points at the repository, on the reasoning that the end of a successful install is the moment a user is most likely to act on it.
0.4.2 - 2026-09-05
Added
--help/-hand--version/-von all three entry points, the gate included (#16, #17).
Changed
- A non-terminal stdin with the flags present and no
--yesstill uses the defaults, and now says so on stderr and in the answers-source line (#18). With flags missing it refuses, as before.
0.4.1 - 2026-09-05
Added
- On npm.
npx agent-personalizeris the install path; 0.4.0 was the first version published (by hand, from the tag). From 0.4.1 every version is published bypublish.ymlwith provenance through npm trusted publishing, so the package page shows where each version was built.npx github:aunysillyme/agent-personalizerstill works.
Changed
- The installer's usage text and rerun hint say
npx agent-personalizer; the README quick start, privacy line and cache note follow.
0.4.0 - 2026-09-05
Two independent audits of v0.3.0 the same day: agy as code reviewer (91/100; it could not execute commands under its own settings) and Codex as a first-time user running the ten-step walkthrough in a fresh temp directory. Every finding was reproduced before its fix.
Changed
signature: nois now coherent across the whole installed tree: the signature rule file is not installed,rules/README.mdloses its row, and the four note templates lose theirLast edited by:line (Codex user seat; check 61 covers the tree).- The installer's rerun hint prints the command that actually ran: the
npx github:form when it ran from npm's cache, else the localnode .../bin/agent-personalizer.jspath (Codex user seat). - The ChatGPT render says, above each fence, to copy only the text inside it (Codex user seat; check 64).
- CI matrix now runs Node 18, 20 and 22 on both OSes, so the
engines: >=18claim is tested (agy). - README: level 1 is "One profile" and states its footprint (about a dozen small files); the privacy line distinguishes the tool (no network) from the
npxfetch that precedes it; macOS/Linux stated, Windows via WSL or Git Bash; a no-sudorecovery when npm's cache is unwritable (npm_config_cache=./.npm-cache). - SECURITY.md said "60-check harness"; it is 72 (agy). CONTRIBUTING names where a new notes tool goes (agy).
Added
--helpand--versionon the installer;--helpon the renderer.--quick: the seven questions that change behaviour (name, tone, length, notes tool and path, write policy, always-ask), the rest default. Interactive only.--answers -reads the answers JSON from stdin, so the scripted path runs in CI.render.cjs --strict: exit 1 and write nothing when a ChatGPT box is over its budget.--rules <dir>names the rule source explicitly.- The ChatGPT target also writes
chatgpt-box1.txtandchatgpt-box2.txt, the plain paste files, covered by--check. - A Windows smoke job in CI (installer, renderer, gate, ESM install through Node);
.gitattributespins LF. .github/workflows/publish.yml: npm publish with provenance and trusted publishing, manual and dormant until the package exists on npm (RELEASING.md).- A harness check that the quoted check count matches the number of checks, everywhere it is quoted.
- Upgrading a level-1 folder to level 3 repoints the home file's installer-written rule pointers at
rules/(Codex round 24).
Changed
- Level 1 is three files. USER.md, AGENT_ONBOARDING.md and the home file(s); the renderer reads the rules from the package when the folder has no
rules/, so nothing is copied.rules/arrives at level 3, where it is yours to edit. The home file's rule pointers say "the rendered block below" until then. The default read-first list isUSER.md, AGENT_ONBOARDING.md. - Every rule's universal block ends with one plain-language "In practice:" line, for the stated audience.
AGENT_ONBOARDING.mdlost its "Standing habits" section (the rules and USER.md already carry it)..agent-personalizer.jsonstores only the answers that differ from the defaults, plus the six pinned safety answers (write policy, always-ask, off-limits, notes tool and path, signature) which are always stored so a future default can never move them (Codex round 24);validate()fills the rest.- The
TOOLtable inrender/onboarding.cjsis the single source for notes tools: the interview option, the kind and the prose derive from it.
0.3.0 - 2026-09-05
Thirteen issues (#3 to #15) filed against 5972b32 by an independent installation-and-behaviour audit, every one reproduced against HEAD before its fix. One fix per issue; the harness grew from 60 to 72 checks, one regression check per issue.
Changed
- Answers drive the rules (#3). A rule can declare
requires: <answer>=<value>in its frontmatter and then renders, and enters the contract, only when the stored answer matches;rules/40-sign-every-edit.mdrequiressignature=yes, and the installer drops its pointer line from the home files onsignature=no.rules/50-output-style.mdno longer hard-codes one shape (bullets, one-line corrections); it defers to the onboarding answers and states neutral defaults for when none are given.rules/00-read-the-profile-first.mdno longer targets ChatGPT or the plain system prompt, which cannot open files. - Re-running with changed answers (#4) regenerates
AGENT_ONBOARDING.mdand every rendered block, and regeneratesUSER.mdonly when it still equals the render of the previous answers byte for byte; an editedUSER.mdis kept and the installer names the changed answers as a conflict. README and paste guide now describe exactly that. - Runtime files are
.cjs(#5):render/render.cjs,render/onboarding.cjs,check/gate.cjs, so an installed copy runs inside a"type": "module"project. The hook, templates, docs and harness follow. - ChatGPT render is budgeted (#6). With answers, Box 1 is a compact profile and Box 2 is the onboarding block (always-ask, off-limits, write policy, then style) plus the
inject: truerules; each box prints its count against the configured limit; an over-budget box is written in full and flagged, never cut; the file says where the remaining rules go (a ChatGPT Project). Without answers the old full render remains as the fallback and says so. notes_pathis honoured (#7): the level-2 scaffold and theCLAUDE.md/AGENTS.mdpointers use it for disk tools; a cloud tool gets no local notes folder and a one-line pointer to the onboarding file; read-only and unknown tools use thenotes/fallback. For disk tools the path is validated as a plain relative path (no.., no leading/). The default read-first list no longer hard-codesnotes/README.md; the onboarding file adds the notes index from the answer.- The session-start contract carries the write policy (#8), and is ordered by consequence: always-ask, off-limits, writes and location, read order, then how to talk.
- Installed files reference only what the destination has (#12): public, version-pinned doc links (
DOCSinrender/onboarding.cjs, checked againstpackage.jsonby the harness) replacedocs/...paths; directnodecommands replacenpm runinhooks/README.mdandrules/README.md. - Docs say what the tool does (#13, #14, #15). README: a "what this does, and does not do" paragraph; level 2 described as instructions the AI follows, level 3 as the automation, level 4 as reading material; "wins at decision time" and "cannot decay" softened to what injection can promise; the gate described as a forbidden-string check with named limits.
docs/tiers.md: the stories are one setup's experience, not measurements; the pointers rule allows a generated, drift-checked block and forbids hand-maintained copies.docs/paste-guide.md: the ChatGPT budget and overflow strategy.
Fixed
- Gate: symlinked parent directories (#9). A working-tree path whose parent is a symlink is not read (it lies outside the scan root); the skip is counted and printed; the index blob is still scanned.
- Gate: forbidden terms in file names (#10). Every shipped path is scanned as well as its text; a path hit is labelled
(path)and can be allow-listed. - Installer preflight (#11). Every path is probed and every existing rendered target is decoded and marker-checked before the first write, so a malformed marker block, invalid UTF-8 or an unwritable target refuses the run with nothing written. Codex round 23 then showed the same gap for sources already in the folder (a kept
rules/*.md, a keptUSER.md): the renderer now exports a source preflight the installer runs first, so those refuse with nothing written too.
0.2.0 - 2026-09-05
Added
- Agent onboarding interview. The installer asks how an AI should work with you (tone, length, mistakes, when unsure, output shape, never-list, read order, where you keep notes, write policy, file naming, signature, off-limits, always-ask-before) and writes
AGENT_ONBOARDING.mdfrom the answers.--answers <json>scripts it;--defaultsor--yesaccepts the defaults.USER.mdis generated from the same answers when absent. Answers live in.agent-personalizer.json, validated on every load. - Notes tool pick-list. Obsidian, Notion, Google Docs, Apple Notes, OneNote, Evernote, Logseq, a plain folder, or other (named). The rendered write section branches by kind: disk tools get filesystem rules, cloud tools get connector rules and an explicit "no filesystem writes", read-only tools get a local fallback folder, other asks before its first write.
- Companions.
docs/companions.md: obsidian-tc (optional,obsidian_tcquestion) and Sierra Catalina's Context Layer, each mapped to the answers it enforces. Per-tool connector table. - Paste guide.
docs/paste-guide.md: which rendered file each AI reads and how, and the re-paste rule. - CI.
.github/workflows/harness.ymlruns the harness and the rollback fault cases on every push and pull request, Ubuntu and macOS. The privacy gate stays local by design and shows as a skip. - Session-start contract carries a short onboarding block when the target allows personal content.
Changed
- Markdown-active answers are backslash-escaped in prose contexts, so an answer of
~~~or<!--cannot restructure the generated file (audit round 17). .agent-personalizer.jsonis validated before the installer's first write, and on every renderer load including--contract(rounds 17, 18).othernames its tool; Apple Notes is described as a separately installed local MCP; obsidian-tc is optional in the render, not mandatory (rounds 20, 21).- File modes in the harness are read with Node; GNU
stat -fis filesystem status, which the first CI run caught.
Fixed
- Cloud tools were handed filesystem write rules and paths named after the workspace (round 20).
- The
othercontract line said read-only;USER.mdpointed read-only tools at an external index (round 21).
0.1.0 - 2026-09-04
Added
USER.mdprofile template with the four firmness rungs;CLAUDE.mdandAGENTS.mdpointer templates.- Seven rules in the three-fence format (
universal/personal/binding:<ai>), each with its origin story. render/render.js: rendersCLAUDE.md,AGENTS.md,GEMINI.md, ChatGPT boxes and a plain system prompt from one source; byte-preserving marker splice;--checkfor drift;--contractfor session-start hooks.check/gate.js: the zero-personal-data gate over exactly what git would ship, against a gitignored list; fails closed.bin/agent-personalizer.js:npxinstaller by AI and level; never overwrites a file it did not create.- Claude Code session-start hook and the plain-prompt fallback.
docs/tiers.md: the five tiers and the failure each prevents.- A worked example (an invented illustrator, one week of notes).
- Harness
test/run.shandtest/rollback.test.js.
Changed, across sixteen audit rounds before the first public commit
- Renderer: target paths validated under the real project root, symlinks refused everywhere, CommonMark-correct fence tracking, invalid UTF-8 refused, staged writes with backup and rollback, generated output re-parsed before staging (rounds 1 to 9).
- Gate: scans index blobs plus differing working-tree copies, symlink target text, every enclosing repository with literal pathspecs; refuses a tracked or symlinked list, bare repos, malformed
.gitmetadata;--allwalks everything but.git(rounds 7 to 15). - Installer: safe destination resolution, strict options, duplicate
--airefused,--dircreated one level at a time. - Harness: exact exit codes, adversarial fixtures, fault injection for the rollback path.