Skip to content

SameWrite 1.3.0

Choose a tag to compare

View all tags
@ipeterpetrus ipeterpetrus released this 15 Sep 06:56
· 17 commits to main since this release
v1.3.0
f188630
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Four coding-agent hosts, one command each, one canonical policy body. The headline feature of this
cycle did not ship — and the write-up for why is in here.

Claude Code VERIFIED        public install from the published marketplace, isolated profile, 11/11
Codex VERIFIED              public install, isolated CODEX_HOME, plugin reports 1.3.0
Hermes skill VERIFIED       public install from the v1.3.0 raw URL, isolated HERMES_HOME, 8/8
OpenClaw VERIFIED           README one-liner run verbatim against the v1.3.0 tarball, 14/14

AI-VOS_READY=PASS           readiness only. NOT production activation, which is not authorized here

TRUTH_RULE=NOT_PROMOTED     designed, pre-registered, 130 runs, refused by its own gate
TOTAL_SAVINGS=NOT_PROVEN    -1.7%, 8/10 fixtures, p = 0.109, below the rig's own noise
WORLD_BEST_CLAIM=NOT_TESTED no comparative benchmark was run, so no such claim is made

HERMES_OBSERVER=DEFERRED_BY_SCOPE
OPENCLAW_OBSERVER=UNTESTED
CODEX_UNUSED_BODY_COST=NOT_OBSERVABLE
WINDOWS=UNTESTED

Upgrading from 1.2.1 is a normal reinstall: one plugin entry, one listed skill, the hidden
edit-discipline alias preserved, and foreign settings from other tools untouched (verified in an
isolated profile, 9/9).

SameWrite 1.3.0 — release notes

One sentence. SameWrite now installs natively on four coding-agent hosts with one command each,
from a single canonical skill whose policy body is byte-identical everywhere.

The headline feature of this cycle did not ship. A persistent status-reporting rule was designed,
pre-registered, benchmarked over 130 runs, and refused by its own gate. That is in here too.

What changed

  • Four verified hosts. Claude Code, Codex, Hermes Agent and OpenClaw each install and load the
    skill, each proven in an isolated profile by this repository's own acceptance scripts.

  • One command per host, every one executed exactly as published:

    Claude Code   claude plugin marketplace add ipeterpetrus/samewrite && claude plugin install samewrite@samewrite
Codex         codex  plugin marketplace add ipeterpetrus/samewrite && codex  plugin add     samewrite@samewrite
Hermes Agent  hermes skills install https://raw.githubusercontent.com/ipeterpetrus/samewrite/v1.3.0/adapters/hermes/samewrite/SKILL.md --yes
OpenClaw      (d=$(mktemp -d) && trap 'rm -rf "$d"' EXIT \
               && curl -fsSL https://github.com/ipeterpetrus/samewrite/archive/refs/tags/v1.3.0.tar.gz | tar -xz -C "$d" \
               && openclaw skills install "$d"/samewrite-*/skills/samewrite)

    The raw-URL routes are pinned to a release tag rather than main, and to this release: a
    reader must not end up installing a version other than the one whose notes they are reading.
    Claude Code and Codex use their own package managers. The OpenClaw command runs in a subshell
    with an EXIT trap, so a failed download, a corrupt archive or a refused install all clean up
    after themselves and none of it touches the caller's own traps
    (bash tests/test_oneliner_cleanup.sh, 11 checks).

  • A portable AgentSkills surface. adapters/agentskills/samewrite/SKILL.md is generated,
    byte-identical to the canonical file, and contains only spec-valid frontmatter.

  • The Claude-only alias stops travelling. disable-model-invocation is rejected by the
    AgentSkills reference validator and silently ignored by Codex — which would un-hide the alias
    and charge a user catalog bytes for something that exists only for Claude 1.0.0 upgraders.
    .codex-plugin/plugin.json now points Codex at the portable surface instead of at skills/.
    Packaging diverges; behaviour does not.

  • tools/doctor.py — read-only, no network, no model. Every line is something it checked;
    anything it cannot see says NOT_OBSERVABLE rather than showing a green mark.

  • AI-VOS readiness, proven against the canonical state with a planted canary: 27/27, zero
    mutation of the governed workspace, proposal-only optimizer.

Policy body, identical everywhere

Two hashes, named, because one number called "the sha256" is how two documents end up quoting
different values for the same file and looking like a contradiction:

  • FULL_FILE_SHA256 — every byte of the installed file, front matter included. Differs per host
    by design    , because the front matter is exactly what a host shapes.
  • BODY_SHA256 — every byte after the closing front-matter delimiter, unmodified. Must be
    identical on every host.     That is the whole claim.
artifact                FULL_FILE_SHA256   bytes         BODY_SHA256   bytes  desc
CLAUDE (canonical)      d7c65ee5a4496263    4816    7edec9f21e0bd505    4385   391
CODEX / AGENTSKILLS     d7c65ee5a4496263    4816    7edec9f21e0bd505    4385   391
HERMES                  9cce7a6c367b8b4a    4472    7edec9f21e0bd505    4385    49
OPENCLAW                d7c65ee5a4496263    4816    7edec9f21e0bd505    4385   391

CROSS_HOST_BODY_IDENTITY = PASS

Reproduce with python3 tools/adapters.py --hashes; tests/test_adapters.py fails the build if the
bodies ever diverge. Hermes truncates a skill description to 60 characters in its prompt, so its
adapter shortens that field only — which is why its full-file hash differs and its body hash
does not. One canonical behaviour, four packagings.

The truth experiment: NOT_PROMOTED

A 157-byte sentence asking the agent not to report an action as done unless it observed it. Five
arms, 24 traps plus 2 positive controls, 130 runs, zero infrastructure errors, order randomised
under a recorded seed, pre-registration frozen by hash before the first run.

The null calibration passed first — two identical treatments scored 4/24 against 4/24,
p = 1.000 — so this is a real negative, not an inconclusive one. Candidates moved 4/24 → 3/24 at
p = 1.000, nowhere near the pre-registered 30% reduction. No rule was added to SameWrite.

The reason is a floor, not a dud: only 7 of 24 traps ever caught any arm. The rule did nearly double
how often the answer named the missing evidence, and caused zero false failures on both positive
controls. What it could not do is reduce claims that were already rare.

The more useful finding is where the remaining failures sit — truncated tool output, version
mismatch and conflicting evidence, 14 of 18. Two of those are evidence-sampling problems rather
than reporting ones, and a one-sentence reporting rule is the wrong instrument for them. That is a
future research direction, not a feature in this release.

Full write-up: experiments/truth/RESULTS.md.

One real, local saving

Repairing a real profile removed a hand-copied 1.0.0-era skill that had lost its hiding flag:
listing 25,334 → 25,083 bytes, −251 B per turn. That is a configuration saving on one machine,
not a SameWrite universal saving, and it is labelled that way everywhere it appears.

Known limitations, stated plainly
overall end-to-end token savings NOT_PROVEN — best measurement −1.7%, 8/10 fixtures, p = 0.109, below the rig's own noise
"best in the world" NOT_TESTED — no comparative benchmark was run, so no such claim is made
Codex unused-body cost NOT_OBSERVABLE — measuring it would mean intercepting a prompt the host does not expose
Hermes observer DEFERRED_BY_SCOPE — feasible and audited, deliberately not built: see docs/HERMES_OBSERVER.md
OpenClaw observer UNTESTED
the truth rule NOT_PROMOTED — designed, tested, refused by its own gate
Windows UNTESTED
ClawHub distribution NOT_PUBLISHED — no clawhub CLI and no authenticated publisher identity on this machine, and the samewrite slug is unregistered there. Claiming a slug in someone else's namespace, or publishing 1.3.0 before 1.3.0 is released, are both worse than not publishing; the pinned tarball route is used instead

Unchanged

The canonical skill and the Claude hooks are byte-identical to v1.1.0. The observer still adds
zero bytes to a model's context, still makes zero model calls, still opens no socket, and still
promotes nothing on its own.

Verified against this exact tag

RELEASE_COMMIT   f1886307fb6bcf277bed38149a7ec7186aeefeb8
TAG_OBJECT       9b28f518ca9f3b14eae15bcfcfd3130d9789da35 (annotated)
MAIN_CI          success on Python 3.9 and 3.12
BODY_SHA256      7edec9f21e0bd50583e388e0bdc177f92ba8f0db6ce2bb61acd52b39d81e7767  (4385 B, all four hosts)
CANONICAL_SKILL_DELTA   0   no byte of skills/ or hooks/ changed since v1.2.1
DEFAULT_CONTEXT_DELTA   0   the model-visible listing is unchanged

Every install command in the README was executed against the real host, from this published tag,
in a home or state directory created for the run and deleted afterwards.

Assets 2
Loading