Intent: the maintainer's answer to the switch plan (ollama post 36efc394ff): start with the bq2 inside the folder, with a bwrap example around it; the repositories bound rather than held in the folder, after a clear one-time move once the processes stop; every app deployed with no step repeated by hand; and no client restart for now. A switch must not lose an app or a repository (ollama post 6a87b6383e).

Expression: ocx-rootfs/README.md's "Start" binds four folders of one DATA: repos, mirror, releases (the release root, which holds OCX's and every app's releases and settings) and letsencrypt. It starts the server with SERVER/usr/bin/bq2, and gives the same command inside bwrap, which shows bq2 only the folder and DATA. Since apps live in the bound release root, every later folder starts with all of them. ocx-rootfs/move-data.cr moves what a folder still holds into DATA, once, by rename. It says what it would do unless given --apply, refuses while a process runs in the folder or where DATA already has the place, and leaves an empty folder for the bind. ocx-rootfs/compare-serve.cr gives the old and new folders the same copy of DATA. It starts the old one as today and the new one through bwrap and its own bq2, asks every repository, CGI program and app of each, and passes only when every answer matches. ocx-rootfs/switch.md holds the steps: build, an HTTPS check, a list of changes made by hand, copies, the comparison, then the switch at the maintainer's time, whose way back is the old folder with the same binds. knowledge-map.json lists it.

Demonstrates: ocx-rootfs/move-data.cr, run on a made-up folder, listed its moves without --apply, then moved the repositories, release root and certificate by rename, leaving empty folders. A second run found nothing to move, and a run whose DATA already held repos refused, exit 1. ocx-rootfs/compare-serve.cr compiles (crystal build --no-codegen ocx-rootfs/compare-serve.cr). Inside this host's bwrap, which showed only the F5 seed's folder (plus host libraries, since that seed's bq2 is not yet static), the folder's own bq2 entered it with sysroot-namespace, and its git listed the Tiny Tapeout repository's refs over HTTPS. crystal run ci.cr: one error, spec/ocx/daemon_pages_spec.cr:253, Fossil's database-locked failure of ticket 00817df826, which passes when rerun. Not yet shown: a server build from the static-bq2 seed, and compare-serve on copies, both on the host.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U Leaf check-in: 39dd9b5c32 user: x86-1-session tags: connector-oauth

Intent: the server's certificate expires on Nov 11, 2026 at 12:32 PM Eastern, and the maintainer wants it renewed from within the root folder, which has no certbot (ollama post a7b8799e93). This is the client ticket 1267e38643 designed, kept small.

Expression: ocx-rootfs/acme.cr is an ACME client (RFC 8555) using only the folder's Crystal and its openssl (LibreSSL). It makes or reuses a P-256 account key, signs each request ES256 through `openssl dgst`, orders the one name, and answers the HTTP-01 challenge on port 80 itself, since Fossil serves only 443. It makes a new P-256 key and a request naming the host in its subject and its one alternative name, then writes privkey.pem and fullchain.pem into OUT, each written beside and renamed into place, the old ones kept as .previous. The contact email is optional. The program prints no key. Fossil reads its certificate only when it starts, so the server is restarted afterwards.

Demonstrates: ocx-rootfs/acme.cr, built inside seed bq2-seed-x86_64-2e1e7bda1b with its Crystal 1.19.1, ran at 1:27 AM Eastern on the server's own machine against Let's Encrypt's staging directory for ollama.openbeagle.org. The name was validated over port 80 from outside, and the staging certificate came back: CN and SAN ollama.openbeagle.org, issuer "(STAGING) Artificial Amaranth YE1", valid to Jan 14, 2027, its public key matching privkey.pem. Fossil 2.27 kept serving the old certificate after its files were replaced, so a restart is needed. crystal run ci.cr had one error, spec/ocx/daemon_pages_spec.cr:253, Fossil's database-locked failure of ticket 00817df826, which passes when rerun; it took 285 s against its 180 s limit, under the host's load.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 08492ab5ea user: x86-1-session tags: connector-oauth

Intent: the maintainer's condition for moving the server to a new folder: "I don't want to restart without apps and without all of the repositories" (ollama post 6a87b6383e). Repositories and git exports are bound from outside the folder, but the release root, /var/lib/ocx/agents-daemon, which holds OCX's releases and every app's releases and settings, is inside it, so a new folder would start without them.

Expression: ocx-rootfs/compare-serve.cr, run in a sealed network and process list on copies, carries the old folder's release root into the new folder, as the switch will, then starts each folder in turn on the same copies of the repositories. It asks each repository's timeline and OCX status, every CGI program in the old folder on each repository, and every app's page and preview found in the old release root, and compares each answer's status and content type. It ends PASS only when every answer matches the old server's, and it notes pages the old server already fails. ocx-rootfs/README.md gains "The server switch, on the host": build, check HTTPS with git ls-remote, list changes made by hand with compare-rootfs.cr, copy, compare, and then the switch at the maintainer's time, whose way back is the untouched old folder.

Demonstrates: ocx-rootfs/compare-serve.cr compiles (crystal build --no-codegen ocx-rootfs/compare-serve.cr) and is formatted. crystal run ci.cr is green: 713 examples, 0 failures, coverage 100.0%. The comparison itself runs on the host, outside both folders, which this session cannot reach; its output goes on ollama post 6a87b6383e.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: a4a9920889 user: x86-1-session tags: connector-oauth

Intent: let the server's imports clone over HTTPS. Every import from an https:// address failed in 0.1 s with "git in the import's sandbox exited 128: Empty enumerable" (ollama posts 40da32abd3 and 77682186eb), because the server's git fell back to the seed's helper, an applet of bq2 that fails at its first request. A step toward the switch the maintainer asked for, which must keep every app and repository (ollama post 6a87b6383e).

Expression: ocx-rootfs/build-rootfs.cr builds the archive-tool links and git-https, git 2.45.2's git-remote-http linked against curl 8.17.0, for both roles, in the order the client's build already passed with; before, only the client got them. The server's /usr/bin/curl becomes the real curl as well, as the client's did.

Demonstrates: ocx-rootfs/build-rootfs.cr compiles (crystal build --no-codegen ocx-rootfs/build-rootfs.cr) and passes crystal tool format --check. Inside seed bq2-seed-x86_64-0b99a242bf, entered with bq2 sysroot-namespace, the git-https step's commands, run by hand from the pinned curl and git sources (both SHA-256 checked), built git-remote-http. With it on GIT_EXEC_PATH, git clone --bare of https://github.com/jadonk/tt-jadonk-tiny-tapeout exited 0, main at 88dd49f, where the seed's own helper printed "Empty enumerable" and exited 128. A whole server build is not run here: its host-only inputs are on the host. crystal run ci.cr had one error, spec/ocx/thread_spec.cr:51, Fossil's database-locked failure of ticket 00817df826, which passes when rerun; and it took 228 s against its 180 s limit, under the host's load.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: d7aecc0a95 user: x86-1-session tags: connector-oauth

Intent: open the README with why OCX is part of the program, as the maintainer confirmed on the Big Picture page and run-194d2a3d asked on ollama post 0b894e6082, shaped like BQ2's and Glint's READMEs and written to docs/writing.md.

Expression: README.md now opens with what OCX is, what it's for, what it's made of (Fossil, Crystal and Linux, replacing the building-blocks sections), how it's shaped (runs, patterns, nodes, tools, releases), what success looks like (six items, each with its status and evidence), and the four virtues in brief. The full axioms and virtues follow, retitled. "What OCX is, in four questions" is folded into the new sections, its links kept. Each claim was checked against the release code (src/ocx/apps.cr, src/ocx/git.cr, docs/protocol-v2.md on connector-oauth) and PLAN.md. Two cold reads by fresh agents with no context were taken; the words they had to guess (contract, clones, daemon, staged, BQ2 seed, client, the branch the server runs) are now defined in place.

Demonstrates: crystal run ci.cr in this trunk checkout passed every guard. Its one error was spec/ocx/run_pages_spec.cr:219, the known random lock failure (ticket 00817df826), which passes when rerun alone. After the second cold read's fixes, spec/guards/documents_spec.cr, size_spec.cr, delineation_spec.cr and derivation_spec.cr pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U Leaf check-in: ed5888b071 user: x86-1-session tags: trunk

Intent: lock in the maintainer's new axiom as a clear proposal, to be tested against experience, with precepts OCX can hold itself to (ocx-console posts 0d2541697a and 4000082c63, relayed by run-194d2a3d on ollama post c9e8a3a7d5).

Expression: README.md adds axiom 6, marked proposed, with its two consequences (every extra syntax is a barrier; many eyes outlast one) and one line on how it relates to axiom 4, Entropy and drift. PLAN.md adds two proposed precepts, One Syntax and Shared Foundations, outside the precept table, since a precept enters the table only with a guard spec; each names the guard it could have. PLAN.md also starts "Testing the axiom" with two cases: Glint's "Words used here" glossary and Glint's choice of Linux and Zephyr. knowledge-map.json lists the new PLAN.md topics, and docs/knowledge-map.md is regenerated.

Demonstrates: crystal run ci.cr in this trunk checkout passed every guard, spec/guards/documents_spec.cr and spec/guards/ledger_spec.cr included. Its one error was spec/ocx/run_pages_spec.cr:219, the known random lock failure (ticket 00817df826), which passes when rerun alone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 2b9add595b user: x86-1-session tags: trunk

Intent: give every project one guide to writing documents a newcomer can follow, in service of the first virtue, Clarity, as the maintainer approved as an approach (ocx-console post 7607cc4e76) and run-194d2a3d asked on ollama post c68bcb0b5d.

Expression: docs/writing.md is the guide as posted, unchanged; its one link already works. README.md links it beside the virtues; knowledge-map.json lists it under "The words and the map", and docs/knowledge-map.md is regenerated from it.

Demonstrates: crystal run ci.cr is green in this trunk checkout (599 examples, 0 failures, coverage 100.0%), including spec/guards/documents_spec.cr, which renders every document and its links.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: cc71cf7c36 user: x86-1-session tags: trunk

Intent: keep trunk's features that OCX lacks in docs/parity-v1.md, as the ticket sort asks (ollama post 8f6c9afbb7), so their tickets can close without the features being forgotten.

Expression: docs/parity-v1.md's "Repositories and OCX itself" table gains two rows read from trunk before the fork (c63c8afd39): each project's Crystal docs served with a standard-library index (ticket fbcaa9a84e), and semantic search from a local model's embeddings. Both are held: the maintainer chose Fossil's full-text search first and has not decided on docs serving (ocx-console docs/ollama-planning.md).

Demonstrates: the guards, crystal spec spec/guards, ran 114 examples with 0 failures, spec/guards/documents_spec.cr among them, which reads docs/parity-v1.md with the rows in it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 0e2d18fbea user: x86-1-session tags: connector-oauth

Intent: record the maintainer's goal that an imported project keep tracking its upstream, "we need to eventually close the circle" (relayed in ollama post 8b49a696b7), placed after the work it depends on (ollama post 0c3e24924f), without scheduling or building it.

Expression: ROADMAP.md's Fossil section gains a paragraph: what project's import does today, the goal (incremental import from the kept marks beside the export, and serving source dependencies such as BQ2's from Fossil), what it would take, that it comes after the release route of ticket 8986816c0b, that its place among the rest is on the Status wiki page, and that BQ2's sources as unversioned files do not wait on it.

Demonstrates: the guards, crystal spec spec/guards, ran 114 examples with 0 failures. crystal run ci.cr failed only spec/ocx/questions_spec.cr:246, on Fossil's "database is locked" of ticket 00817df826, and that example passes when run again. The import's marks are named in src/ocx/commands/project.cr (GIT_MARKS, FOSSIL_MARKS).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 0f2d19483f user: x86-1-session tags: connector-oauth

Intent: let a clean build of x86-1's client from seed bq2-seed-x86_64-0b99a242bf pass the archive-tools step, which stopped with "Error creating symlink: 'llvm-ar' -> '/usr/bin/ar': File exists" when the host session built it (ollama post bab4fb063e).

Expression: ocx-rootfs/build-rootfs.cr's link_archive_tools makes /usr/bin/ar and /usr/bin/ranlib point at their llvm- tools only where no such link is there yet, as the libc links above it already do. This seed ships both as llvm-ar and llvm-ranlib.

Demonstrates: ocx-rootfs/build-rootfs.cr compiles (crystal build --no-codegen ocx-rootfs/build-rootfs.cr). The seed, unpacked, holds /usr/bin/ar -> llvm-ar and /usr/bin/ranlib -> llvm-ranlib. The host session's build with this change resumed past the step, and llvm-config --targets-built in the folder printed X86 ARM AArch64 (ollama post bab4fb063e). crystal run ci.cr ran 713 examples with one failure and one error: the check-in guard on af38773c64, whose Demonstrates line named the build program without its folder and is now amended, after which spec/guards/checkin_spec.cr passes; and spec/ocx/run_pages_spec.cr:222, Fossil's database-locked failure of ticket 00817df826, which passes when run again.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: cda5d59c5b user: x86-1-session tags: connector-oauth

Intent: let x86-1's client be rebuilt from BQ2's seed whose LLVM targets X86, ARM and AArch64, which Glint needs to build for ARM, as the maintainer asked (ocx-console post 688e1532c9, relayed in ollama post b28a74c516).

Expression: ocx-rootfs/build-rootfs.cr's SEED names bq2-seed-x86_64-0b99a242bf.tar.gz, sha256 0c93201ba563033f8100a436e9e7a6283b931263c4b753f8cf5e5c5d11d122c4, as BQ2 reported it (bootstrap-qcow2 post b337765b1c). The seed is not on the server yet, so the builder puts it in DOWNLOADS under that name; fetch refuses any other content.

Demonstrates: ocx-rootfs/build-rootfs.cr compiles (crystal build --no-codegen ocx-rootfs/build-rootfs.cr). crystal run ci.cr ran 713 examples with 0 failures, but went over its 180 s limit (262 s) while the host's load was 17 on 8 CPUs. The build from this seed is shown when it is run on x86-1's host.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: af38773c64 user: x86-1-session tags: connector-oauth

Intent: let an agent's save run OCX's checks before it commits, as the maintainer approved (ocx-console post fe5ad9e6e8). The first Gemma run's ocx_test answered "the project names no tests", so its save committed with none run (ticket 29062db5de).

Expression: ocx-tests.json at the root names one test, ci, as crystal run ci.cr, which ocx_test runs and save requires to pass.

Demonstrates: crystal run ci.cr in a trunk checkout with the file is green: 599 examples, 0 failures, coverage 100.0%. The next Gemma run's ocx_test on ticket 29062db5de shows it run by the agent.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: a043a3fba0 user: x86-1-session tags: trunk

Intent: let a login an admin approved be asked again for other letters with one more press, as the maintainer expected register to do (ocx-console post 13dfe73616). Collect refused a code other than the approved one with a text register took as a failure, so no link was printed (bootstrap-qcow2 post 774683323c).

Expression: collect now refuses another code as "approved for another code, and for this one it is not approved yet", so register prints the approve link for its new code. Approve already sets a login it made with exactly the letters asked and a new password. docs/invite.md says to ask for the whole set of letters and that the old password stops working. docs/interfaces.md says the same, and the done row leaves docs/backlog.md. The admin's press stays the only protection, as it is for a first ask.

Demonstrates: spec/ocx/logins_spec.cr, "registers from the asker's own machine": from a new file, the link asks goi24; once pressed, the login holds exactly goi24 with a new password, and the old one no longer signs in. crystal run ci.cr is green: 713 examples, 0 failures, coverage 100.0%.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 05ad811298 user: x86-1-session tags: connector-oauth

Intent: the maintainer decided that a share needs no name rule: "if it is in uv, it should be replaceable ... anyone with y can write it—that is our one and only protection from it being overwritten in real life. don't produce machinery that can't work." (ocx-console post 7085f98c83, relayed in ollama forum post 98286416a9). So each architecture's BQ2 seed can be one file, replaced at each build.

Expression: Bounds.wider no longer refuses a share whose name lacks $$run$$; its other rules stand, ends_with_parent among them. HISTORY.md decision 133 amends decisions 107 and 132; protocol v2's refusal 1 drops the clause. The rule itself is written into Parsimony in README.md, with no guard spec, since none can check it (ollama forum post b4cb02763c: the maintainer repeated "don't make machinery that can't work" of a spec for it).

Demonstrates: spec/ocx/decisions_spec.cr: a run from R alone sharing under builds/$$run$$/x, ocx2-release.tar, seed/x86_64.qcow2 or $$process_user$$run$$ starts. crystal run ci.cr: 713 examples, coverage 100.0%; its three failures, two the known rcvfrom lock (ticket 00817df826) and one a compiler's stdin read, pass rerun.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: afd3df339f user: x86-1-session tags: connector-oauth

Intent: listing a clone's unversioned files ran fossil uv sync first, so every clone that listed them pulled every file, each ~100 MB BQ2 seed among them (ollama forum post 5bdae9f13d, the maintainer's ask relayed by the One Place session, from ocx-console a85cc28929).

Expression: ListUnversioned reads the clone's remote and lists it without a sync: over https or http, Fossil's /juvlist page, unsigned, any user or password in the remote left out, as the receipt tool is asked, read within Exchange's time and at most LISTED bytes; a file remote by reading that repository; a clone with no remote, its own files. A remote over neither, a server not reached, or a page that is not the list is refused, saying why. unversioned_get and unversioned_put still sync, as they did. docs/remote-agents.md says the list fetches no file.

Demonstrates: spec/ocx/checkout_tools_spec.cr: a file added on a served repository is listed by name and size while the clone holds no unversioned file; an ssh remote and a dead server are each refused, naming why; the file-remote edit test lists as before. spec/ocx/dry_run_spec.cr: the list's dry run runs to its end with no remote. crystal run ci.cr: 713 examples, coverage 100.0%, its one error the known rcvfrom lock (ticket 00817df826), whose spec passes rerun.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 1724d0f5e2 user: x86-1-session tags: connector-oauth

Intent: the maintainer asked that a staged app be seen before it is accepted: "Accept should not be required before preview" and "once it is staged, a preview should be available" (ocx-console posts 9d5d8673f4 and 498cf64f4d, relayed by the One Place session). ext/ocx2-cgi-next/app/APP ran the app's current release, so a staged console could not be previewed.

Expression: Apps.serve takes the preview explicitly, under precept C6 (Explicit Arguments), from the context the program's name sets, and Apps.program picks the app root's staged/next for the preview and current otherwise, as Projects.program does for OCX's own program. With nothing offered, the preview's 404 says no release is offered as next. The tools doc and the modules list say so.

Demonstrates: spec/ocx/apps_spec.cr: before accept, the page answers that nothing is accepted while the preview runs the staged build with Fossil's request; the preview's 404 names next. crystal run ci.cr: 713 examples, coverage 100.0%, its one error the known rcvfrom lock (ticket 00817df826), whose spec passes rerun.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: ba21098522 user: x86-1-session tags: connector-oauth

Intent: the maintainer's first try at adding the connector in Claude (8:51 PM Nov 5, record.jsonl on ollama; screenshot ocx-console 4b8cbe87ce) opened authorize in Claude's in-app browser, signed out of Fossil. The page's sign-in link sent them to Fossil's login, which after sign-in showed "Not Found: Page not found: https:". Fossil 2.27's login sends the browser back to the repository's URL, a slash and g (`%R/%s`, its src/login.c), and the page gave g as an absolute https:// URL.

Expression: the consent page gives g as the path below the repository, ext/ocx2-cgi/oauth/authorize?FIELDS, which Fossil's rule turns back into the same authorize request.

Demonstrates: spec/ocx/connector_spec.cr: signed out, authorize's sign-in link carries g such that SITE, a slash and g is the authorize request it answered. End to end against a local `fossil server --nojail --repolist --extroot --notfound --files` started as hosts.json starts the public one: signed out, authorize asks to sign in; its link opens Fossil's login, which keeps g; the login POST is sent back to exactly the authorize request, which offers approve; approve 302 to https://claude.ai/api/mcp/auth_callback with the state; token 200; initialize 200; notifications/initialized 202; tools/list, five tools; ocx_status as voice. The guards pass, with the known database-is-locked flake (ticket 00817df826) and the fixture teardown race rerun green on their own; coverage 100.0%.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 73894cf022 user: x86-1-session tags: connector-oauth

Intent: x86-1's daemon, started again at 8:59 PM with its output kept, ended after one round with "stopped: the server answered 400 to /chat-poll" (ollama post 904dc7dc97): it named neither the project nor the request, nor why Fossil refused it; x86-1 asked that the stop name both. Fossil's chat-poll page never answers 400 itself (its src/chat.c); a 400 comes from its request parsing, malformed_request in src/cgi.c, which says why in its answer's first line.

Expression: Server::Asked#request refuses an answer that is no success or redirect saying the server's URL, the method and the path with its query, and the answer's first QUOTED (200) bytes on one line. Neither holds a secret: the URL takes no user part (Asked.reachable), the sign-in sends its password in the body, and the cookie is never shown.

Demonstrates: spec/ocx/daemon_spec.cr: chat-poll answering 400 "Bad HTTP Request: malformed HTTP header" raises from wait, naming the server, GET /chat-poll?name=-1 and that line, and daemon.log's last line says the same. spec/ocx/daemon_refusals_spec.cr, spec/ocx/daemon_pages_spec.cr, spec/ocx/receipt_spec.cr and the guards pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 197a37c0d8 user: x86-1-session tags: connector-oauth

Intent: x86-1's daemon, started on 0397ade722 after its follow, did one round on every clone at 8:33:54 PM and then exited with no line in daemon.log saying why (ollama post 55dabc25d3): Daemon.serve ends the daemon once any clone's fiber ends, and a fiber ends when its round or its chat wait raises; a round logged only a refusal, and a wait logged nothing, its exception reaching only standard error, which that start discarded. The maintainer: OCX fails visibly and logs why (their Nov 4 direction on watchdogs).

Expression: Served#round logs any exception that ends it, a refusal by its message and any other by its class and message (Daemon.said), and Served#wait logs a failed chat wait as a round that read nothing, "the chat wait failed: ...", each then raised as before, so the daemon still ends.

Demonstrates: spec/ocx/daemon_spec.cr: chat-poll answering an error item twice, as Fossil does for a login it no longer signs in, raises from wait, and daemon.log's last line says "the chat wait failed: chat-poll answered an error after signing lab-1 in again: chat is off". spec/ocx/daemon_refusals_spec.cr, spec/ocx/messages_spec.cr and the guards pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: dbaa1241eb user: x86-1-session tags: connector-oauth

Intent: the console's stage run a90d00d294 kept its files unsent at 6:49 PM, x86-1 holding no y on ocx-console; the app press gave x86-1 y, ocx-console's receipt said ghjorwyz4C, and yet x86-1 answered the publish at 8:05 PM "Publication file 1 failed: the server does not allow unversioned files" (ocx-console 31b6758dd3). Server::Asked keeps each login's letters for as long as it lives, and a daemon keeps one for each clone for as long as it runs, so the "no y" it read before the press stood.

Expression: Server::Asked#forget_capabilities forgets the letters it was told; Daemon::Served#round calls it as each round begins, so within a round a login is asked once, as before, and a letter given since is seen at the next.

Demonstrates: spec/ocx/daemon_refusals_spec.cr: against a stand-in receipt tool whose answer for x86-1 gains y, the letters are asked once, read again only once forgotten, and then allow sharing. spec/ocx/daemon_spec.cr and spec/ocx/publication_spec.cr pass, as do the guards.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 735f5f9bbc user: x86-1-session tags: connector-oauth

Intent: the console's stage run a90d00d294 ended 0 with its publications pending, as x86-1 held no y on ocx-console; once it did, the publish that sends them again could come only from a person's reply on Fossil's forum page, as `ocx2 post` refuses OCX messages by design and Fossil has no forum command (ocx-console post 3379b5ba9e). docs/backlog.md keeps an item until the maintainer picks it.

Expression: docs/backlog.md gains the row, with what to work through first: who may send it, whether a tool of its own or a kind post takes, and that it writes only an empty publish to the run's clone login.

Demonstrates: spec/guards/documents_spec.cr and the other guards pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 86e6bf656f user: x86-1-session tags: connector-oauth

Intent: the maintainer (ocx-console 627c50a1c5): accept's answer said "changes: 0397ade7223b..0397ade7223b", current..next once next was accepted and both named it, where it means previous..current, ff56aef644c9..0397ade7223b; the accept of ff56aef644 read the same. Their hint, "it is likely due to handling the rollback": the range was read from the links after the tool rewrote them, and a rollback leaves next as it was, so it read forward where it went back.

Expression: ReleaseState takes `replaced`, the release current named before the tool moved it, read before the move: accept, rollback and follow pass it, and changes reads replaced..current; otherwise, while another release is offered, current..next, as before. Its help says both.

Demonstrates: spec/ocx/release_commands_spec.cr: a second release staged reads first..second; accepted, first..second, not second..second; rolled back, second..first. spec/ocx/follow_spec.cr, spec/ocx/release_import_spec.cr, spec/ocx/styles_spec.cr and the guards pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 0f32e7b2f5 user: x86-1-session tags: connector-oauth

Intent: the coverage ratchet of c62e565500's CI: three statements of the connector no spec ran. src/ocx/connector_oauth.cr's refusal of an approval naming no key could not be reached, as the request's own check refused it first; src/ocx/connector_store.cr's skipping of a config row that does not read as a connector, and its refusal of a value signed by the key that is not of the kind asked, a registration traded as a code.

Expression: approve finds the key named first and refuses at once where there is none, then checks the request. spec/ocx/connector_spec.cr presses approve naming no key, trades a client_id as a code, and keeps a connector row of 'not json'.

Demonstrates: spec/ocx/connector_spec.cr: approve for id 0000 says no key is named 0000; the client_id traded as a code is 400; the store lists voice alone beside the broken row. spec/coverage/run.cr over it misses no statement of src/ocx/connector.cr, src/ocx/connector_oauth.cr, src/ocx/connector_store.cr or src/ocx/commands/connector.cr.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 0397ade722 user: x86-1-session tags: connector-oauth

Intent: the release plan says the builder's watcher "sends the files as the builder's login, which holds y" (ocx-rootfs/README.md, "The steps", step 1; ticket 8986816c0b: the trusted entry's reporting login gets y). The console's stage run a90d00d294 ended 0 with its files unsent because x86-1 held no y on ocx-console, and 1bb37a2b79 answered that with a tool of its own, built without reading the plan. The maintainer decided (ocx-console card e362fab960, decision 07768e08eb): "Yes, app gives its builder y, and drop the separate tool."

Expression: app refuses a builder the served repository has no user for (ReadOwnCapabilities), keeps the builders row, then adds y (AppCommand::SENDS_FILES) to the builder's own letters on the repository's user table by AddOwnCapability's one statement, every other letter as it is then, never through Fossil's user page, which saves a user across its login group. write_unversioned, its spec and its entries in the tool lists, the design, the survey and docs/protocol-v2-tools.md go. The command order puts send_runs before app, which now requires it.

Demonstrates: spec/ocx/apps_spec.cr: x86-1's ghjorwz4C becomes ghjorwz4Cy, its login group's other member keeping ghijorwyz4C, a second press changing nothing, a builder the repository lacks refused with the row as it was. spec/ocx/command_spec.cr, spec/ocx/mcp_spec.cr, spec/ocx/dry_run_spec.cr, spec/ocx/styles_spec.cr and the guards pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: c62e565500 user: x86-1-session tags: connector-oauth

Intent: the full CI of 1bb37a2b79 failed five specs. spec/ocx/styles_spec.cr: the list offered accept and rollback where the release offered was not whole, or was already served, as the optional --app the merge gave them made every card with a field to fill offer its form, the plan made with no input judging none. spec/ocx/mcp_spec.cr: MCP lists write_unversioned, as it lists send_runs.

Expression: Page::Cards#runs? offers a form without judging it only when the tool has a required field the viewer fills in (fills?, by the input's required list); otherwise it follows the dry run of what was given, when the page shows one beside the card (judged, from Cgi.page), and else the plan made with no input, which is the form left empty. So OCX's accept is offered only when it would run, and accept?app=console, a link naming the app, offers its press once that dry run runs to its end. spec/ocx/mcp_spec.cr lists ocx_write_unversioned.

Demonstrates: spec/ocx/styles_spec.cr and spec/ocx/mcp_spec.cr pass again; spec/ocx/apps_spec.cr: with console's release staged and nothing offered for OCX, the page of accept?app=console offers its press and the page of accept offers none.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 6175490b42 user: x86-1-session tags: connector-oauth

Intent: one release carrying both, to be staged and accepted together (the tracker's walk-through for the maintainer: import, accept, the y press for x86-1 on ocx-console, the app press, the console's import and accept). The console's stage run a90d00d294 ended exit 0 with its three files unsent: x86-1 holds ghjorwz4C on ocx-console, no y, and Fossil passes over unversioned files a user may not write and says nothing. The maintainer decided (ocx-console card d9513c5b14, decision 62c0fcd144): "Add a one-login y press to the release being staged, then press it for x86-1." Fossil's own user page saves a user across its login group, so saving x86-1 on ocx-console would set its letters on ollama, where it holds ghijorwyz4C, taking its i away there.

Expression: web-install (5532f86776) merged, its conflicts resolved to both sides: the connector's and the app's routes ahead of the page's form, both config row prefixes (Fossil::CONNECTOR_KEYS, Fossil::APP_BUILDERS) and their C2 exemptions, the tool lists and dry-run samples, and the design, facts and survey entries of each. write_unversioned, needing admin as send_runs does, refuses a user the served repository lacks and Fossil's own users, then adds y to the user's own letters on the served repository's user table by AddOwnCapability's one statement, every other letter as it is then, and answers what it may do as receipt reads it.

Demonstrates: spec/ocx/write_unversioned_spec.cr: y added to x86-1's ghjorwz4C, another member of its login group keeping ghijorwyz4C, a second press changing nothing, and each refusal. spec/ocx/connector_spec.cr, spec/ocx/apps_spec.cr, spec/ocx/command_spec.cr, spec/ocx/dry_run_spec.cr and the guards pass. Through a local Fossil 2.27 started as scripts/hosts.json starts the public one, this build as ext/ocx2-cgi: the connector's whole flow, as in 756cf18f86, and app/console 404 with none accepted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 1bb37a2b79 user: x86-1-session tags: connector-oauth

Intent: the console, and later other projects' programs, installed with clicks and no shell (ollama forum post a224b383c5), at the address the maintainer chose, /ocx-console/ext/ocx2-cgi/app/console (ocx-console 8fdbe7d537), on its own branch from the accepted release ff56aef644 (ocx-console 7a15514592). The server already runs ext/ocx2-cgi as the accepted ocx2 (scripts/confine.cr, its CGI mode), so ocx2 runs the project's accepted program; nothing changes in scripts/, and nothing is written on the server but the release itself (the maintainer, ollama forum post e173a3d120).

Expression: app, a setup login's press, keeps the stage pattern and builder whose runs may be releases of the served repository's program APP in a config row, Fossil::APP_BUILDERS, named in docs/design/facts.json and spec/guards/config_state_spec.cr. Importing takes the program's name: its three files, its smoke ("usage: APP"), and its builders, OCX's from ROOT/settings/builders.json, an app's from that row (Importing.builders_of). import, accept and rollback given --app act on ROOT/apps/REPO/APP (OCX.root!), which import makes first. src/ocx/apps.cr names that root and runs ROOT/apps/REPO/APP/current/APP for ext/ocx2-cgi/app/APP and under it, with the request fields OCX's wrapper read as its whole environment and the body on its standard input, its output the answer; 404 where none is accepted; ext/ocx2-cgi/app alone stays the press's page. Its one Process call is listed in spec/guards/external_programs_spec.cr.

Demonstrates: spec/ocx/apps_spec.cr: the press, each refusal, the row read back; a builder's run of console imported as OCX's is, by receipts, into its own root, OCX's root untouched, accepted, and run by the page with its path, viewer and body; rollback with none previous; 404s; --app outside a served repository. spec/ocx/release_import_spec.cr, with spec/support/release_import.cr taking a program, spec/ocx/dry_run_spec.cr, spec/ocx/command_spec.cr, spec/ocx/release_commands_spec.cr and the guards pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U Leaf check-in: 5532f86776 user: x86-1-session tags: web-install

Intent: the maintainer, on the connector writing /var/lib/fossil/.well-known and a directory beside the repository (ollama forum posts a3372955ce, e173a3d120 and afbe369331): serve the metadata from a URL the Fossil repository provides; the server's config is in the repository, in scripts/hosts.json, and nothing is written to the server's file system unless it must be, a change there meaning bubblewrap and a restart.

Expression: a key's SHA-256 and who approved it are a config row a person's press sets, Fossil::CONNECTOR_KEYS (ocx-connector-oauth-LOGIN), as register's approvals are (Fossil::REGISTER_KEYS), named in docs/design/facts.json and spec/guards/config_state_spec.cr. The issuer is the repository's uv/oauth/ID, and connector keeps its metadata as the unversioned file oauth/ID/.well-known/openid-configuration, the address OpenID Connect's discovery appends to an issuer, which Fossil serves at /uv/. A registration's client_id and a code are their claims signed with HMAC-SHA256 by the key's kept SHA-256, so nobody keeps them; a code is good until it expires, ten minutes; a token is kept nowhere, as no later call can show it to the page. The record of requests is the unversioned file oauth/record.jsonl, its newest hundred lines, written as the viewer; oauth/record goes. .ocx-connector/ and the .well-known file and its writability check go.

Demonstrates: spec/ocx/connector_spec.cr: the whole flow, with the metadata read from the unversioned file, nothing made beside the repository, no key in it or the record; each authorize mistake; a code by another grant type, a wrong verifier, unsigned or for another client refused; the press never replacing a key unasked, revoking key and metadata. Through a local Fossil 2.27 started as scripts/hosts.json starts the public one, --acme --nojail --repolist --extroot --notfound to an HTML page --files "*.ico,*.html": the 401; the resource metadata; at the two root addresses a client tries first, /.well-known/oauth-authorization-server/... and /.well-known/openid-configuration/..., the --notfound HTML page, 200; at the repository's /uv/oauth/ID/.well-known/openid-configuration the metadata, JSON, as application/x-fossil-artifact; register 201; authorize asking to sign in, then offering approve; approve 302 to claude.ai's callback with the state; token 200; initialize 200; notifications/initialized 202; tools/list, five tools; ocx_status as voice; the record as /uv/oauth/record.jsonl; and the server's directory holding only the repository and its page.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 756cf18f86 user: x86-1-session tags: connector-oauth

Intent: the full CI of 4ba274c5eb was green but for the coverage ratchet: src/ocx/connector.cr's answer to a request Fossil names no repository for, and src/ocx/connector_oauth.cr's sweep of expired codes, which runs only once a key holds a code, were run by no spec.

Expression: spec/ocx/connector_spec.cr approves a key a second time, which keeps its first approval and gives another code, and asks the connector with no FOSSIL_REPOSITORY, answered 404.

Demonstrates: spec/ocx/connector_spec.cr; spec/coverage/run.cr over it leaves no statement of src/ocx/connector.cr, src/ocx/connector_oauth.cr, src/ocx/connector_store.cr or src/ocx/commands/connector.cr missed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: e3bbd9aa1e user: x86-1-session tags: connector-oauth

Intent: the maintainer asked for OAuth inside the CGI, the key in the URL naming the login, with no relay and no Fossil rebuild (ocx-console posts e43387a603 and 9e1f19d0fa; plan in ollama forum post ebf37502be), tested end to end against what claude.ai requests, with a record of each OAuth request reaching the page. Fossil 2.27 hands an ext program no Authorization header and serves no path holding /. under ext/, so neither the token nor OAuth's well-known addresses can reach the CGI; but `fossil server --acme`, as the public server runs, serves any file under the repositories' .well-known (its src/main.c).

Expression: connector, a setup login's press, keeps the SHA-256 of a new key for a login holding neither admin nor setup, off, beside the served repository in .ocx-connector/ (Connector::Store, no config row, precept C2 (Single Owner)), and writes its issuer's RFC 8414 metadata at .well-known/oauth-authorization-server/ plus the issuer's path; it refuses where that cannot be written, naming the one host step. On the page (src/ocx/connector.cr, connector_oauth.cr): a call with a key not yet approved is 401 with WWW-Authenticate naming oauth/resource?id=ID (RFC 9728); oauth/register (RFC 7591); oauth/authorize, which asks the viewer to sign in, and offers a setup login an approve press, taken only as a POST from OCX's own page, which turns the key on and redirects with a code; oauth/token trades that code once with PKCE S256 (RFC 7636), keeping only the token's hash. An approved key's POST is one MCP message as the login, listing and calling only status, questions, thread, post and ticket (Mcp.handle's `only`). Every request is recorded in record.jsonl with only non-secret field values, shown at oauth/record to a setup login. The wrapper passes CONTENT_TYPE and HTTP_USER_AGENT.

Demonstrates: spec/ocx/connector_spec.cr: the whole flow; each authorize mistake; approval refused to voice and to a press from elsewhere; a code traded once, by its grant type and verifier; 405, 401, 404 and 503; the press never replacing a key unasked, revoking key and metadata, refusing admin, bad names, unknown logins and an unwritable .well-known. spec/ocx/dry_run_spec.cr and spec/ocx/harness_spec.cr pass, as do the guards. And through a local Fossil 2.27 `fossil server --acme --nojail --repolist --extroot`, as hosts.json starts the public one, with this build as ext/ocx2-cgi, a client asking as claude.ai does (MCP authorization 2025-06-18): initialize 401 → resource metadata → issuer metadata by RFC 8414 path insertion, application/json → register 201 → authorize signed out asks to sign in, signed in offers approve → approve 302 to https://claude.ai/api/mcp/auth_callback with the state → token 200, the same code again 400 → initialize 200 → notifications/initialized 202 → tools/list, five tools → tools/call ocx_status as voice; the record listed each request, holding no key, code, verifier, state or token, and Fossil served neither .ocx-connector/ nor a directory under .well-known.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 4ba274c5eb user: x86-1-session tags: connector-oauth

Intent: the maintainer asked to revert the key-in-URL connector (ollama forum post 3979474437, "please revert the crap") and, on its stalled CI, to start development on a new branch from the release the server runs, ff56aef644 (ocx-console post 859db5d886). protocol-v2-daemon is left as it is, abandoned. Four of its pieces work without the connector: the daemon's whole-or-missing clone (4b74d9d5ff); post asking only the rights a write needs (63ee550c91, Codex review finding 3); the ticket tool the maintainer's plan asks for, "Write (gated): create a Fossil ticket" (c0c3e91f68), page-only under precept A5 (Outcome Home); and the backlog row for register re-asking an existing login (d186f9c9bc).

Expression: those check-ins and post's format fix (bb925dffe8) cherry-picked onto ff56aef644, their conflicts resolved to the files as protocol-v2-daemon held them once the connector (9690013918, abe85d676f and its specs in c531fb9ade) was taken out: no connector command, route, key rows or MCP tool filter; ticket's comments name the connector as not built; the ticket specs c531fb9ade added are kept.

Demonstrates: spec/ocx/ticket_spec.cr, spec/ocx/post_spec.cr, spec/ocx/clone_server_spec.cr, spec/ocx/command_spec.cr, spec/ocx/dry_run_spec.cr and spec/ocx/mcp_spec.cr pass (39 examples), as do the guards (114); the whole of spec/ocx on protocol-v2-daemon with the connector removed, the same files, passed: 561 examples, 0 failures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 636255fc8f user: x86-1-session tags: connector-oauth

Intent: Claude's "Couldn't start sign-in" at 2:01 PM, 2026-11-05 (ollama forum post 3979474437): voice's key had been replaced, as any second press of connector, or a reload of the page after one, made a new key and turned the old off; and Fossil 2.27 sent the connector's text/plain 401 wrapped in its HTML page, and its 202 and 405, which named no type, as application/x-fossil-artifact.

Expression: Connector.answer heads every answer, 202, 401 and 405 among them, with Content-Type: application/json, each but the 202 a JSON body. connector refuses a login that has a key already unless given replace, which then turns the old one off. docs/protocol-v2-tools.md says both.

Demonstrates: spec/ocx/connector_spec.cr, each answer's status and type, a second press refused with the first key still acting, and replace turning it off. Through a local Fossil 2.27 server with this build as ocx2-cgi: initialize 200, notifications/initialized 202, tools/list, a ticket call, GET 405 and a bad key 401, each application/json. crystal run ci.cr -- unpaced: green, 704 examples, 0 failures, coverage 100.0%. Leaf check-in: abe85d676f user: x86-1-session tags: protocol-v2-daemon

Intent: the maintainer wants capability changes to come as a register approve link (ocx-console post 6d67d547ab), since Fossil's Users page drops R when it saves a user, but register refuses a login already approved (bootstrap-qcow2 post 774683323c). And docs/backlog.md keeps an item only until the maintainer picks it: Claude's connector was picked (ocx-console 093cd88820) and is built in 9690013918 and after.

Expression: docs/backlog.md gains the row for register re-asking an existing login, with what to work through first, and loses the connector's row.

Demonstrates: spec/guards/documents_spec.cr and the other guards pass. check-in: d186f9c9bc user: x86-1-session tags: protocol-v2-daemon

Intent: the stage build of bb925dffe8 (run 38445d23fe) broke ci.cr's coverage ratchet: three statements no spec ran, connector's refusal of a login that is not one, the connector's key given empty, and ticket's refusal of empty text.

Expression: spec/ocx/connector_spec.cr presses connector for a login of other letters and calls mcp with key empty; spec/ocx/ticket_spec.cr files a ticket with no text. No code changes.

Demonstrates: spec/ocx/connector_spec.cr and spec/ocx/ticket_spec.cr pass; crystal run ci.cr -- unpaced on the client is green, 704 examples, 0 failures, coverage 100.0%, 2 executables build. check-in: c531fb9ade user: x86-1-session tags: protocol-v2-daemon

Intent: the stage build of c0c3e91f68 (run e71d724ea2) failed ci.cr's format check on src/ocx/commands/post.cr, so no release could be offered.

Expression: one line of src/ocx/commands/post.cr, NEEDS's alignment, as crystal tool format writes it; nothing else changes.

Demonstrates: crystal tool format --check src spec tools ci.cr reports nothing, and spec/ocx/post_spec.cr passes. check-in: bb925dffe8 user: x86-1-session tags: protocol-v2-daemon

Intent: a voice chat, or a person on the page, puts an idea into Fossil as a ticket, as the maintainer's plan for Claude's connector asks first ("Write (gated): create a Fossil ticket"); no tool did, since create sends a run and post only changes a ticket that exists (Codex review finding 6, ollama forum post 23f89f73da).

Expression: the tool ticket starts an Open ticket, titled and described as given, as the caller on the served repository, through Fossil::Place#new_ticket, Fossil's own ticket change artifact with a new 40-digit id. It refuses a clone's checkout, where a run works, so precept A5 (Outcome Home) holds, and its guard names the files as a person's prompt; it refuses text holding an OCX message, a title not one line, a caller without NewTkt (n), and, where the server holds new tickets for a moderator, one without ModTkt (q). The connector lists and calls it.

Demonstrates: spec/ocx/ticket_spec.cr, a ticket filed on the page read back with its title, status, description and sender, and each refusal, a clone's checkout among them. check-in: c0c3e91f68 user: x86-1-session tags: protocol-v2-daemon

Intent: post required Fossil's Check-In (i) for every reply and ticket change, so a login holding only forum and ticket rights, as the voice connector's does, could not reply on a thread without being given code commits (Codex review finding 3, ollama forum post 23f89f73da).

Expression: post needs any of i34wc to be offered; once it knows where it writes, a reply needs WrForum (3) or WrTForum (4), and a change to an existing ticket WrTkt (w) or ApndTkt (c), never NewTkt (n), with Check-In (i) still making either. The connector now lists and calls post too. docs/protocol-v2-tools.md says both.

Demonstrates: spec/ocx/post_spec.cr, a login holding ojrn234 replying on the page and refused a ticket change, and one holding ow refused a reply and making the ticket change. check-in: 63ee550c91 user: x86-1-session tags: protocol-v2-daemon

Intent: Claude's voice client reaches OCX while the maintainer travels, as one login holding only what it needs and never admin or setup, its key made by a person's press, never by an agent or by voice (ocx-console posts 093cd88820 and 7fdfd4fd79; ollama forum posts f101f1c518, 2527feb6cd and 2be6a86560).

Expression: the page's ext/ocx2-cgi/mcp takes one JSON-RPC message a POST and answers it as MCP's, as the login whose key's SHA-256 the served repository keeps (Fossil::CONNECTOR_KEYS, an unsynced config row a person sets, exempt from precept C2 (Single Owner) as secrets and approvals are), with the capabilities Fossil gives that login as each call arrives, listing and calling only status, questions and thread; 401 for a key acting as no one or as a login now holding admin or setup, 405 for anything but a POST. The setup-only tool connector, withheld from MCP, makes the key and shows its URL once, or revokes it. The key is in the URL because Fossil 2.27 passes an ext program no Authorization header.

Demonstrates: spec/ocx/connector_spec.cr, a press making the URL, the key kept only as its hash, tools listed and called only from the connector's set as voice, a login given setup refused, a replaced, revoked or unknown key refused, a notification and a GET, and an agent refused. check-in: 9690013918 user: x86-1-session tags: protocol-v2-daemon

Intent: a clone stopped part way, as beaglebadge's social-badge clone stopped at 217 KB (ocx-console post a16a7762a1), was left at the clone's own file, which the daemon takes for a whole clone, so it never cloned again and the clone had no server-code to send runs with.

Expression: CloneServer clones into NAME.fossil.part beside the clone, removing any left there, and renames it to the clone only once Fossil succeeds; docs/protocol-v2-tools.md says so in the daemon's bullet.

Demonstrates: spec/ocx/clone_server_spec.cr, a failed clone leaving no clone and a leftover part replaced by a whole clone with its server-code. check-in: 4b74d9d5ff user: x86-1-session tags: protocol-v2-daemon

Intent: the maintainer wants sessions to know the names they give conversations with ocx-console's menu (ocx-console 5c026f2e95, through its session); the console writes the name as the raw tag console-name on the thread's first post.

Expression: Place#conversation reads that tag on the root as the repository holds it now, a cancelled one as none (Fossil::CONVERSATION). Acting.delivery adds "conversation" to a post's line between from and text, absent when the thread has no name; thread's answer gives it for a thread, through the check read_conversation. docs/protocol-v2.md and docs/protocol-v2-tools.md say so.

Demonstrates: spec/ocx/acting_spec.cr (case 12) gives a post before naming without the field and one after with it, non-ASCII kept; spec/ocx/thread_spec.cr reads none, then the name, then none once cancelled. crystal run ci.cr green on its second run (the first failed a spec under load).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: ff56aef644 user: x86-1-session tags: protocol-v2-daemon

Intent: x86-1's dry run of follow (2026-11-05) stopped with "no answer in full within 10.0 seconds" fetching ocx2, which is 30 MB and a third more as base64.

Expression: Server::Asked#release passes FILE_WITHIN, ten minutes, when it asks for a file; the listing keeps Exchange::WITHIN and its byte bound.

Demonstrates: spec/ocx/follow_spec.cr still green; crystal run ci.cr green.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 91121abe61 user: x86-1-session tags: protocol-v2-daemon

Intent: the maintainer's pull of stage run 4fd8667247 on the server stopped with "Regex match error: UTF-8 error": every fossil command's output passes through Fossil.masked, and uv cat of a built ocx2 is not UTF-8.

Expression: Fossil.masked scrubs bytes that are not UTF-8 before masking; what a command printed is still returned unchanged. The import fixture's program now holds such bytes, as a built one does.

Demonstrates: spec/ocx/release_import_spec.cr failed with the server's error before the fix and passes after, through accept with exact bytes; crystal run ci.cr green.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 11f8fa7483 user: x86-1-session tags: protocol-v2-daemon

Intent: the maintainer (2026-11-05) wants to pull the candidate on the server the ordinary way, run inside the server folder; the installed ocx2 predates import, so the candidate''s own ocx2 does it.

Expression: ocx-rootfs/pull-candidate.cr RUN SETUP-USER writes settings/builders.json if absent (the published stage pattern, the run''s builder and clone), exports the run''s ocx2 from ollama''s unversioned files, and runs its import and accept on /var/lib/ocx/agents-daemon, keeping the release before as previous. scripts/ is unchanged since the live release 34ed26f294, so the server''s host programs stay.

Demonstrates: ocx-rootfs/pull-candidate.cr, run by hand on a copy of this clone and a fake root it wrote builders.json and ran the candidate''s import, which refused with no counted started, as a clone holds no server receipts; the whole import needs the server''s own repository. crystal spec spec/guards/: 0 failures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: e87708e831 user: x86-1-session tags: protocol-v2-daemon

Intent: the maintainer (post e7d3f56b13) would rather stage everything and copy the first release in than rehearse it first; so the hand install the README listed is one command, which says what it would change, then does it.

Expression: ocx-rootfs/install-first.cr SERVER CLIENT REPOSITORIES RUN [--apply] gathers the stage run''s files from the server''s ollama.fossil and builds the host programs before either root changes; with --apply it refuses while the client''s daemon holds node/daemon.lock, keeps each replaced host program as PROGRAM.previous and the replaced release as previous, and writes settings/builders.json. README, "Before the first ordinary release", now names it; test-rootfs.cr is a rehearsal when wanted.

Demonstrates: ocx-rootfs/install-first.cr, run by hand on fake server and client roots, with a copy of ollama.fossil and stage run 4fd8667247: the dry run changed nothing, --apply refused while the lock was held, then installed 8ed4852d40a2 on both, current and previous as expected, ocx2 answered its usage, and a second dry run found all six host programs unchanged. crystal spec spec/guards/: 114 examples, 0 failures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 64530c70ab user: x86-1-session tags: protocol-v2-daemon

Intent: stage runs of check-ins before 0388e725f3 hold CI's pace limit and crossed it on x86-1 (23529c7fc5 at 183 s, 9df11d18f0 at 194 s). Stage run 528f437c3e, of 0388e725f3 itself, ended exit 0 at 181 s with its note "not held for a release build", and its three files are on the server with matching hashes. A host session following "The rehearsal, on the host" would otherwise pick an older check-in and fail. Clarity: the procedure names the floor. Parsimony: one sentence.

Expression: ocx-rootfs/README.md, "The rehearsal, on the host", step 1 says both stage runs must be at or after 0388e725f3, and why.

Demonstrates: spec/guards/documents_spec.cr, spec/guards/delineation_spec.cr and spec/guards/size_spec.cr pass, with crystal spec spec/guards/ at 0 failures. The full ci.cr was not run for this change to Markdown; this check-in is also to be the rehearsal's NEXT release.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 8ed4852d40 user: x86-1-session tags: protocol-v2-daemon

Intent: the maintainer, 2026-11-04, answering whether `ci.cr -- build` may skip the pace limit (post 11c1ece96f): "yes, it can skip the limit for now." Two stage runs of 6390f85333 on x86-1, 23529c7fc5 and 9df11d18f0, failed at 183 s and 194 s against the 180 s of docs/precepts/pace.json. The stage pattern runs CI at nice 19 and ionice idle on a machine shared with sessions, and even a run at normal priority there took 182 s. Clarity: one mode, named for what it leaves out. Loose Coupling: the published stage pattern, sha3 ebf7864f…, does not change. Demonstrability: a spec pins the call. Parsimony: plain CI and its other modes keep the limit.

Expression: ci.cr gains the mode `unpaced`. It runs the plain checks, and a run over the limit is green, its result line noting the time and "not held for a release build". `build OUT ARCH` runs its checks as `crystal run ci.cr -- unpaced`. The usage comment and USAGE.md, "Stage, accept or roll back a release", say so. gate.txt still carries only the examples line, so the note shows in the printed result, which the stage run's report keeps.

Demonstrates: spec/ocx/ci_commands_spec.cr records the checks build asks for and expects "run ci.cr -- unpaced". spec/guards/interfaces_spec.cr's "accepts every mode ci.cr's usage advertises" covers the new mode, and spec/guards/pace_spec.cr still holds. `crystal run ci.cr` on x86-1 is green: 701 examples, 0 failures, coverage 100.0%, 2 executables build.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 0388e725f3 user: x86-1-session tags: protocol-v2-daemon

Intent: ticket a3cf4ee8a7 says `ocx2 post` answered ok for four posts by x86-1, which lacks Fossil's Check-In (i), and blames a refusal Fossil sends as a message, "message pull only - not authorized to push", which `Sync.unsynced` does not match. Before changing `unsynced`, pin what Fossil 2.27 really prints, so the fix rests on a shown fact. Demonstrability first; Parsimony: no code changes until the cause is shown.

Expression: a seventh example in spec/facts/fossil_spec.cr serves a fixture repository, clones it as lab-1, which holds goC4w, and writes a forum post in the clone. A `sync` exits 0, sends nothing and prints no error, so `unsynced` finds nothing. Then, with lab-1 holding goC4wy, as x86-1 does, a `push` exits 0 and prints "Error: not authorized to write", which `unsynced` returns, so `post` refuses. The post never reaches the served repository. docs/fossil.md is at its 400-line limit, so the fact is in the spec and on the ticket only. No OCX path pushes with `sync`, and the ticket's words were not reproduced, so the four posts' cause is still open on the ticket.

Demonstrates: spec/facts/fossil_spec.cr, 8 examples, 0 failures; crystal spec spec/guards/ with 0 failures. The full ci.cr ran green on this branch at 0773efb93c (700 examples, coverage 100.0%), in 182 s at normal priority.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: e6cb9ed39f user: x86-1-session tags: protocol-v2-daemon

Intent: the handoff c333d6ce16 makes the rehearsal, ocx-rootfs/test-rootfs.cr, the next step toward each next ordinary OCX release being deployable and able to be rolled back on the public server and the client. The README's "What shows it works" promised stage, import, accept, follow and rollback, but test-rootfs.cr checked none of them. The maintainer said (a62694c6d0) that a Claude session on the host, outside both roots, may run it. Clarity: one procedure in the README that a host session follows by itself. Loose Coupling: the rehearsal imports builds x86-1 already made, from a copy of the repository, so nothing is built inside the sealed network, which has no shards. Demonstrability: each step is a printed check. Parsimony: no new tool, only two more arguments.

Expression: test-rootfs.cr takes FIRST and NEXT, two ended stage runs of two check-ins. After its existing checks it stops everything it started, builds this checkout's serve, confine and receipt-reader into both folder copies, installs FIRST's files by hand on both roots, and writes the server's settings/builders.json. Then it checks that import refuses an unlisted login, another pattern and a changed file, changing nothing. It checks that NEXT is imported, accepted and listed by /ext/ocx2-cgi/release, that follow moves the client so its daemon ends, and that exactly one daemon then runs NEXT by its executable's inode. Last, it checks that both roll back to FIRST the same way. The README gains "The rehearsal, on the host": two stage runs with --of the full 64-digit hash, paths from the running start commands, copies taken with sqlite3 .backup, and unshare --map-root-user, since the maintainer has no sudo. "What shows it works" now says what the test shows, and that a client which missed a release is not shown.

Demonstrates: crystal build --no-codegen ocx-rootfs/test-rootfs.cr type-checks; spec/guards/documents_spec.cr, spec/guards/delineation_spec.cr and spec/guards/checkin_spec.cr pass, with crystal spec spec/guards/ at 114 examples, 0 failures. On x86-1, unshare without --map-root-user is refused, and with it the port setting and both mounts work. The rehearsal itself has not run: it needs the host. The regular expression for a run's check-in and the server-code query were tried on this clone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TUH8KKBs1b4wzoSCq6zh7U check-in: 0773efb93c user: x86-1-session tags: protocol-v2-daemon

Intent: the maintainer said on 2026-11-04 that docs/invite.md ignored the new tools and should be rewritten, and agreed to the rewrite. Its steps had a person make a login with a one-time password and an agent replace it with curl; register, approve and collect now do that, and no person types or sees a password. Clarity: two ways in, a run first, then the invited agent's six steps and the person's two; Loose Coupling: each step links to the USAGE.md section that owns it rather than restating it; Demonstrability: every step is the one x86-1-session took joining ocx-console on 2026-11-04, recorded in that repository's SETUP.md; Parsimony: the curl recipes for the login and forum pages are gone.

Expression: docs/invite.md rewritten: register asking goi24, collect, clone with the URL on standard input and Fossil's password line filtered, reply on the inviting thread with ocx2 post, audit, close; the person presses the approve link and later clears the capabilities. It states the gap that OCX has no clone step for an agent outside a run, and that ocx2 post writes only replies. AGENTS.md's summary of those steps follows it.

Demonstrates: spec/guards/delineation_spec.cr, spec/guards/checkin_spec.cr; crystal spec spec/guards/ green. check-in: 6390f85333 user: x86-1-session tags: protocol-v2-daemon