Hosted Git compatibility

Hubnot needs an ordinary Git remote to accept, advertise, and transfer objects reachable through refs/hn/*. This document records direct observations only; it does not infer support for providers or configurations that were not tested. Hosting UI/API state is not evidence.

Local bare Git

Automated black-box acceptance creates a fresh local bare remote and uses only ordinary Git transport. The three-run operational scenario and focused hostile fixtures directly establish:

• separate actor and candidate refs are advertised and exact-fetched;
• actor and candidate objects cross the remote without a Hubnot service;
• explicit selections exclude unselected refs;
• valid selections can promote while invalid, mismatched, over-budget, and dependency-missing selections remain absent from accepted refs;
• an identity-free depth-1 clone reconstructs exact policy, identity, candidate, CI, review, decision, merge, and Git commit bindings;
• selected recovery can obtain an exact missing predecessor while the repository remains shallow and without importing an unselected actor.

On 2026-08-31, TestOperationalAgentMemory separately created two actors, recorded all six memory kinds, superseded and retracted exact records, added a cross-actor challenge, and published the two exact memory stream refs through nh sync. A credential-disabled fresh clone initially had no local or accepted memory refs, private identity, or index. After saving only those two full stream selectors, it synchronized through quarantine, rebuilt .git/nh/memory/index-v0.json, and recalled the same full IDs, ownership, anchors, lifecycle edges, evidence, trust classes, content digests, and inert handoff lists. Deleting and rebuilding the index produced byte-identical index and recall JSON. Three consecutive offline runs passed.

That local-bare observation establishes ordinary Git transport and protocol behavior, not public-host policy or retention. Hostile memory containing shell- and tool-shaped prose, ANSI controls, newlines, and Unicode produced no marker effect and remained nested under the recall data boundary.

The detailed run evidence and bounds are in self-hosting-alpha.md.

Local bare hn pre-public observation

On 2026-09-01, candidate hn version 0.0.1-dev, built from 84d82ca52b86b520cbfad7a155f789b0eb914f66, initialized an ephemeral actor, opened one signed issue, and synchronized it to a fresh local bare Git remote. Ordinary git ls-remote advertised exactly:

544b02c42d391502c18b4171071888c79f772731 refs/hn/actors/fd388bbc342c4f157ee72c4dca963bb5f1c4416af3290216b3e33134cbba9497

A fresh clone saved that full actor selector with positive budgets, ran hn sync origin, and promoted the same OID to:

refs/hn/remotes/origin/actors/fd388bbc342c4f157ee72c4dca963bb5f1c4416af3290216b3e33134cbba9497

hn log reconstructed the signed issue. Searches of both the remote advertisement and verifier refs returned no refs/nh/*. The sequence used only git init --bare, hn init, hn issue open, hn sync, hn replication select, git ls-remote, and git for-each-ref over a local file transport. This is pre-public candidate evidence, not a claim that the final cutover commit or refs/hn/* have been published to GitHub.com.

GitHub.com private observation

The 2026-08-29 test used a temporary private repository under an authenticated personal account. GitHub accepted actor refs outside the branch and tag namespaces:

refs/nh/actors/2f2534ef...
refs/nh/actors/e675ea17...

Ordinary git push, git ls-remote, and explicit Git fetch established:

1. a new actor ref was accepted and advertised unchanged;
2. a fresh clone fetched actor history only after Hubnot supplied an explicit custom refspec;
3. a second actor published a separate signed comment history;
4. the first clone fetched it and reconstructed both verified events;
5. every event commit and parent needed by those histories transferred.

That observation preceded candidate refs and selected quarantine, so it does not claim those newer surfaces for the private-repository configuration.

GitHub.com public operational-alpha observation

Before the operational-alpha public mutation on 2026-08-30, ordinary Git advertisement for https://github.com/LynnColeArt/hubnot.git returned:

refs/heads/main e8b0955a3d10b5677fb0dfaba42580f8f4473080
refs/nh/actors/36944394addccd027292abc8183f332af8b4590925291ee8f1e8d8f09446b7dd 2b0e261037e0760f83536fedfad044e95deff1fe
refs/nh/proposals/4fd8813d5849241173eb48bbebd6858535545fa514a80b88d3fdae0e59377290 327199e686dbe0bcc4cc4bf92f10237472b083a8

These values were obtained with git ls-remote, not a provider API. Stage 1 published the two actor histories and policy candidate before separately advancing main to 08291d6f363fa68ca66b84d922d46cc8302fc9a0. Stage 2 advanced the two actor histories and added the release candidate before separately advancing main. The final ordinary-Git advertisement was:

refs/heads/main 7e41574792be828ba507e5df7adda71662475483
refs/nh/actors/36944394addccd027292abc8183f332af8b4590925291ee8f1e8d8f09446b7dd 4d9caba6548c50eba49ed89dc8ee45960226c01e
refs/nh/actors/10425b859ef6c86afe46e9586a1c26a9fe007f7a0aa38a46a3c316fcfc69a1db 8bc318bdd0a8a37e3fc55d692651811124516c14
refs/nh/proposals/4fd8813d5849241173eb48bbebd6858535545fa514a80b88d3fdae0e59377290 327199e686dbe0bcc4cc4bf92f10237472b083a8
refs/nh/proposals/52aee3a8a6fd110321b3506e9408397158d02d3dfe63747cb445effa157e0a95 a3ef855ca062c191a6c7e5c14709de67bdfb4f1c
refs/nh/proposals/122a6dc9ba587f787897395e89c38bd48aef1f68dc5d7dc35121562a1073371d 07b64dfdb11dc376644e068f201ca998bb54defe

A fresh HTTPS clone at depth 1 saved exact selections for the two actors and three proposals with budgets 10,000 events, 100,000 objects, 16,777,216 bytes per object, 1,048,576 bytes per attachment, and 268,435,456 total bytes. It promoted 24 verified events, reconstructed the release as merged with its exact amended policy and role-distinct evidence, and remained shallow. It had no private keyring, legacy identity record, or local public actor ref.

The complete event, policy, object, and commit record is in the public proof record.

GitHub.com public agent-memory observation

On 2026-08-31, the public endpoint https://github.com/LynnColeArt/hubnot.git was exercised using ordinary git and nh commands only. No provider API, UI state, Docker operation, service, model API, copied keyring, or copied index participated.

At 2026-08-31T17:51:34Z, git ls-remote showed main at 7e41574792be828ba507e5df7adda71662475483, with no proof/agent-memory-v0-20260831 branch and no advertised memory refs. The proof then pushed only the exact new branch and exact actor/memory refs. It did not force, delete, wildcard-push, tag, or update main. At 2026-08-31T17:55:59Z, advertisement was:

fc6e3fff308658f3c1e33ab819631d6001ddf9a5 refs/heads/proof/agent-memory-v0-20260831
3f1257bd84ef354390b5f82d25c056663976bc3e refs/nh/actors/9584bdce9dcefc43a05c8dd34c77f4967419d610a28041e2da4b4fad33f726fc
c7700147f216d7edd474b33364a6ea5415c965b5 refs/nh/actors/d91339c192742a9a7676ec2da81d8e6177dbe0b07058729e6c0674a7a20e01f1
a0db69f5dbf77ecaa2f6d3a5ebe0f8e686369d51 refs/nh/memory/9584bdce9dcefc43a05c8dd34c77f4967419d610a28041e2da4b4fad33f726fc/52fa6c89cf8a92cc1be017bd1d49957f0ce0210f56386f7a57af7b9249add648
a50f4a004dcd89858b599d9ed85df7ece72173a3 refs/nh/memory/d91339c192742a9a7676ec2da81d8e6177dbe0b07058729e6c0674a7a20e01f1/aaaaa298cb8ff0bb41a6e59c899fe64e0a6d0989921e2ace7b12359c0c8d77f7

The branch policy digest was sha256:ff545e1bbe0e07176b6ee7639093f1143b17af4797dd3af165bf4d14c19d7804. It qualified both exact actors and all six memory kinds. The author stream contained this decision and handoff:

sha256:3579329703b5ee53d0c124e42fadcacafd9f74eb062f426c96f0ffedff211333 decision
sha256:e6113e0985758ebaaa39f130df7b26f13913fee6bf732c5b034c0c1c8ae94e1f handoff

The successor stream contained challenge sha256:f468dcea0749451404c3e21806e2cbddd438d8cf06c0c6408e7f3a94c18f3c7d against that exact decision, with evidence bound to proof commit fc6e3fff308658f3c1e33ab819631d6001ddf9a5.

Publication of each brand-new local stream completed before its first import snapshot knew the ref, so that first nh sync reported the exact new stream as dependency-missing while still advertising it. After git ls-remote confirmed the full ref and OID, an unchanged retry promoted every exact selection. No accepted ref was inferred from push output.

A new --single-branch HTTPS clone ran with an empty isolated home, GIT_CONFIG_NOSYSTEM=1, terminal prompts and askpass disabled, token variables empty, and no SSH agent. Before selection it had no refs/nh/*, .git/nh identity, index, embedding, adapter, or Docker state. The verifier saved these full selectors with positive budgets:

nh replication select origin \
  --actor 9584bdce9dcefc43a05c8dd34c77f4967419d610a28041e2da4b4fad33f726fc \
  --actor d91339c192742a9a7676ec2da81d8e6177dbe0b07058729e6c0674a7a20e01f1 \
  --memory sha256:52fa6c89cf8a92cc1be017bd1d49957f0ce0210f56386f7a57af7b9249add648 \
  --memory sha256:aaaaa298cb8ff0bb41a6e59c899fe64e0a6d0989921e2ace7b12359c0c8d77f7 \
  --max-events 10000 --max-objects 30000 \
  --max-object-bytes 16777216 --max-attachment-bytes 1048576 \
  --max-total-bytes 134217728
nh sync origin
nh memory index rebuild
nh memory recall --at HEAD --json

All four selections reported promoted. Clone, synchronization, index build, and recall took 1.35 s, 2.70 s, 0.05 s, and 0.05 s respectively on the observed client. The index had mode 0600, 5,805 bytes, source fingerprint sha256:4203a45253c628959b593929bc17bec1f7a4f622c7a22d8a87566ced5cf0df47, and file SHA-256 2b14b5c3d82aaa9f98589e00f6d65fc248c74444501860e5a86723acef6ef04c. Deleting and rebuilding it produced the identical digest.

Recall returned exactly two qualified, applicable, active, evidence-resolved, signature-valid records with query digest sha256:3db4aec9870c5ea2b792b23153e2141e7f3e79cda88129f933276b89368f21a9. The decision retained the full challenge ID, and the handoff retained all four inert lists. A record attempt exited 1 with no identity; the verifier did not gain either originating key.

The proof establishes exact transport and reconstruction for these two actors, streams, branch bytes, client, endpoint, and observation time. It does not establish distinct humans, semantic truth, prompt or operational authority, permanent host retention, portable pre-download quotas, provider generality, moderation, redaction, deletion, or global erasure.

Clone behavior

An ordinary git clone and ordinary branch fetch do not import custom refs, because Git configures a branch refspec by default. This is expected. After an explicit selection, hn sync requests exact selected actor/candidate refs and stores validated roots below refs/hn/remotes/<remote>/*.

The primary branch and collaboration refs are separate namespaces and separate publication outcomes. A host accepting one does not imply that the other was updated; verify both with ordinary Git:

git ls-remote --symref origin HEAD
git ls-remote origin 'refs/heads/main' 'refs/hn/actors/*' 'refs/hn/proposals/*'
git ls-remote origin 'refs/hn/memory/*/*'

Limits not established by these observations

• GitHub Enterprise Server or organization-specific ref policies;
• very large actor/candidate counts, histories, or packs;
• very large memory-stream counts, histories, or packs;
• provider garbage-collection and long-term custom-ref retention;
• deletion, rollback, force-push, and global redaction behavior;
• partial-clone filters;
• portable hard pre-download quotas;
• providers other than the explicitly observed local Git and GitHub.com configurations.

A successful transport observation does not make the host a policy authority, prove separate humans, or strengthen the local runner sandbox.