Phase 5 of docs/agents/GITEA-SETUP.md, plus a correction to Phase 2 that the
runbook could not have known it needed.
gitea-mcp v1.7.0 goes into both images, pinned by the sha256 the release
publishes and smoke-tested with `--version` at build time, so a bad pin fails
the build instead of the agent. Each entrypoint registers it at user scope for
that container's own identity, remove-then-add so a restart is idempotent.
The token is passed BY PATH. `-e GITEA_ACCESS_TOKEN=$(cat …)` would write it in
cleartext into ~/.claude.json, which every session in the container reads;
GITEA_ACCESS_TOKEN_FILE is new in the pinned version and leaves the secret in
its read-only mount. Verified against the binary's own --help, not assumed.
The tool filter stops being an experiment. The names are in the release README:
each agent gets issues, notifications, labels, milestones and pull requests, and
NOT `pull_request_review_write`. That one matters because separate identities
open a hole the runbook did not name: Gitea refuses to let an author approve
their own pull request, and does nothing about sylph-decoder approving
sylph-port's. Two agents could satisfy `required_approvals = 1` between
themselves and then merge, since branch protection blocks pushes to main and
never blocked merges.
Withholding the tool is defence in depth; the controls are in branch protection,
and both docs now say so: approvals whitelisted to the human so an agent's
approval does not count, merges whitelisted to the human so an approved PR is
still merged by a person. Phase 2's check gains the step that actually tests it
-- approve the throwaway PR yourself, then confirm the agent STILL has no merge
button. Without that step, the check passes on an instance where the agents can
merge each other's work.
Also settles two entries on the runbook's own "not verified" list: the tool
filter names, and the Gitea version (1.25.5, whose API schema carries
enable_merge_whitelist and enable_approvals_whitelist under those names).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01McNbzUeq1KRBWs4G6X2YVj
Both agents stopped, and the decoder diagnosed it itself:
"I received '2' and '1' but I don't have a pending question those would
answer -- I was in the middle of setting up the /loop cron job."
claude-autonomous matched the BARE SUBSTRINGS 'Choose', 'trust' and 'accept' to
answer Claude Code's one-time first-run gates. The /loop prompt is echoed into
the terminal, and that day's briefs contain 'accepted as-is' and 'least
trustworthy' -- so expect matched the agent's OWN INSTRUCTIONS and typed 2\r and
1\r into a running session, which then sat waiting for a human to explain them.
The old comment argued a multi-word pattern 'never matches' because the gate
text wraps. True of a literal string, false of a whitespace-tolerant regex, which
is what these now are: \s+ spans the wrap, and the terminal is 200 columns wide.
Measured, old against new, against the real brief text and a real gate:
{accept} brief 0 gate 1 (case-sensitive; briefs say 'accepted')
{Yes,\s*I\s+accept} brief 0 gate 1
{trust} brief 1 <- the trigger
{Do\s+you\s+trust\s+the\s+files} brief 0
Two defences, because one is not enough for something that can type: patterns
prose cannot match, and gates skipped ENTIRELY on resume (SYLPH_SKIP_GATES) --
a resumed session cannot show a first-run gate, so there is nothing to answer
and everything to lose. Timeout cut 90s -> 25s for the same reason.
Also: SYLPH_OWN_LOGIN. Remote Control stopped registering under the long-lived
token, and the likely reason is scope -- `claude auth login` requests
user:sessions:claude_code and the token's auth status reports no email, org or
subscription. A per-agent `claude auth login` restores Remote Control AND avoids
the rotation collision, because each agent holds its own grant rather than a copy
of one. The flag stops the entrypoint seeding the host's credentials over it.
The rotating OAuth credential file is why the agents kept parking, and a
long-lived token removes the failure by construction instead of recovering from
it after the fact.
MEASURED 2026-09-04. ~/.claude/.credentials.json holds a refresh token that
ROTATES ON USE. Seeding both containers from the host left three clients holding
one token; the first to refresh invalidated the other two, and on the failed
refresh Claude Code CLEARS the stored tokens -- writes empty strings, keeps the
metadata, and parks at "Login expired".
decoder credentials emptied 13:04:28
decoder last transcript 13:04:29 <- one second later
The emptying and the park are the same event, which is why it never self-heals:
not a stale token a retry could fix, but no token at all, with no browser in the
container to complete /login. A hollow file passes every "does it exist" check --
508 B healthy against 280 B emptied -- which is how three separate diagnoses
missed it. And recovery re-armed the bug: after re-seeding, host and decoder held
the IDENTICAL refresh token hash.
`claude setup-token` issues a long-lived token against the same Claude
subscription. Checked, not assumed: `claude auth login` defaults to --claudeai
and it is `--console` that means Console/API billing, so this is not the separate
API bill. `CLAUDE_CODE_OAUTH_TOKEN` is recognised by the installed binary.
Passed as an ENVIRONMENT VARIABLE, both halves of the failure are gone: nothing
rotates, so peers cannot invalidate each other, and there is no file for Claude
Code to empty on a failure.
Both launchers read $HOME/.sylph-claude-token if present -- same pattern as
SYLPH_GIT_CREDENTIALS -- and both entrypoints skip OAuth seeding entirely when
the variable is set, because copying the rotating file in would re-create the
exact collision the token exists to remove.
Inert until the file exists. Without it, nothing changes.
Also worth recording for the preflight work: `claude auth status` prints JSON
with loggedIn/authMethod/subscriptionType. That is a far better SessionStart
assertion than checking a file exists, and it would have caught this on the first
iteration rather than the third incident.
The decoder died mid-task and it took four separate findings to explain, each
of which read as something else:
1. OOM-KILLED, REPORTED AS A CLEAN EXIT. `OOMKilled: true` with **ExitCode 0**.
So `--restart on-failure` would treat a memory kill as a successful finish
and leave the agent down -- the policy has to be `unless-stopped`.
2. THE JOB CAP WAS SET AND THEN REMOVED THREE LINES LATER. build-reborn has
always exported CARGO_BUILD_JOBS, but a raw `cargo test --release -p
sylpheed-formats` never reaches the wrapper. Adding `-e CARGO_BUILD_JOBS` to
the launcher did not help either: the entrypoint recomputes and exports over
it unconditionally. An explicit value now wins, and says so in the log.
3. THE MEMORY CONSTANT WAS WRONG. `mem_gib * 2 / 3` assumes ~1.5 GB per job;
release rustc on this workspace needs ~2 GB, and 4 jobs in 6 GB is what died.
Divisor is now 2.
4. `--continue` CANNOT RESUME AN ABRUPT DEATH, which is the only kind we get.
It resolves through ~/.claude.json's per-project `history`/`lastSessionId`,
and MEASURED mid-session both are None -- they are written at a graceful
shutdown. A killed container never writes them, so `--continue` answered
"No conversation found to continue" with 33 MB of transcripts in the volume
beside it. Persisting .claude.json did not help, because the fields were
never populated in the first place; that attempt is removed rather than left
in looking useful.
The TRANSCRIPTS are durable and named by session id, so the entrypoint reads
the id off the newest one for its cwd and passes `--resume <id>`. Verified
on both agents: each reattached to its exact prior session and appended to
the same file rather than opening a new one.
The /loop prompt is still passed alongside `--resume`, so the loop is RE-ARMED
rather than merely restored -- a resumed conversation with no wake-up scheduled
answers once and stops, which looks like resuming and is not.
Restarting into the same death is guarded at the other end: a start less than
120 s after the previous one begins FRESH instead of continuing back into
whatever killed it. That fired correctly during this work.
On resume the agent is told it was restarted, that its in-progress work is
uncommitted in the tree, that any build or capture it had running did not
finish and its absence is not a result, and which wrapper to prefer over a raw
release build.
Credentials were seeded only when the container's copy was MISSING. So when a
session expired, the file still existed, the copy was skipped, and restarting
changed nothing -- the one recovery path a human has, re-logging in on the host,
could not reach the containers at all.
Now re-seeds whenever the host's copy is newer. Newer-wins rather than
always-copy, because a container refreshes its own token mid-run and that copy
may legitimately be the fresher of the two.
Found when both sessions expired: host credentials at 16:30, containers holding
14:20 and 14:24.
Three defects, all mine, found by checking instead of assuming.
**They never exchanged a word.** SendMessage=0, ListAgents=0 across both new
sessions. PROTOCOL.md specified in detail what a message MAY and MAY NOT do and
never said how to send one or that the other agent was addressable -- they knew
that last time only because the human told them directly, and rebuilding with
fresh volumes wiped it. Policy without mechanism is prose. Now documented with
the two addresses, a worked example, and an instruction to introduce themselves
on the first iteration rather than waiting to have a question.
**The decoder lost the disc and the ISO.** They used to arrive inside the project
mount and silently stopped when /work became a clone. Silently is the word: the
disc-gated tests SELF-SKIP without SYLPHEED_DISC and report green, so a whole
test suite would have passed while measuring nothing. Both are now mounted
explicitly, the ISO at a stable path so run-canary does not depend on host
directory names.
**Both agents shared one Claude state directory.** They share the host's
~/.claude, and once both working directories became /work they resolved to the
same projects/-work/ -- two supposedly independent agents writing to one place,
which undoes the point of separate checkouts. Each now has its own volume, seeded
once from the host with credentials only, so a token refresh writes locally and
neither can corrupt the host's auth.
Also widened the pacing rule. It banned ScheduleWakeup by name; the decoder then
scheduled itself an hourly cron job -- not harmful, but the same instinct that
ended a run yesterday, through a door I had left open. Now: no self-scheduling by
any route.
Mount audit after the changes: shared and intentional are the exchange volume and
the read-only credential seed. Everything else -- repo, Claude state, cargo,
target, canary, disc, ISO -- is per agent or one-sided.
Two bugs, both mine, both found by starting the thing.
**Volume mount points must exist AND be owned by the agent before USER agent.**
Docker seeds a named volume from whatever the image has at that path, ownership
included, and creates a ROOT-OWNED directory when the path is absent. Either way
the agent cannot write, and the failure surfaced far from its cause: "clone
FAILED", with no permission error anywhere in sight. The port's own Dockerfile
already carried a comment explaining this trap, which I then walked into for
/work and /exchange.
**The clone guard checked for a .git directory, not a usable HEAD.** A clone
interrupted partway -- the container was removed while one ran -- leaves a .git
with no commits, and a presence check then skips the retry forever and hands the
agent an empty repository that looks like a checkout. It now verifies HEAD, and
clones via a temp directory so a partial result never lands in /work at all.
Also: the port launcher's path defaults still assumed the old repo root, so it
mounted no disc; and the stale /reborn notice is gone now that there is one
repository.
Verified running: both agents cloned c58196b, `share` on PATH from /work/tools,
/exchange agent-owned, canary at /canary for the decoder, disc at /disc for the
port.
The last structural fix for the collision class that has bitten three times. Both
containers now clone the repository into their OWN named volume instead of
bind-mounting a human's working tree, so an agent's local git config cannot
capture a human's commits, a credential helper cannot leak a container-only path
onto the host, and a `git add -A` cannot sweep another party's in-flight files.
Cloned once at startup and never auto-pulled: pulling under a running agent
moves files out from under whatever it is mid-edit, which is the same bug again.
Accepted knowingly: Claude Code keys per-project memory off the working
directory, so moving off the host path starts that memory empty. The corpus in
docs/ is the memory that matters and it travels with the clone.
Other changes:
* docker/agent -> docker/decoder; the launcher is sylph-decoder. Roles, not
"the agent", now that there is more than one.
* /reborn is gone -- one repository now, so the port reads HANDOFF from its own
checkout rather than through a live read-only mount of someone else's tree.
* Canary mounts separately at /canary; it stays a fork tracking upstream.
* A shared `sylpheed-exchange` volume at /exchange, with tools/ on PATH so
`share` is available in both.
* The decoder's credential file gets the .host-copy treatment the port already
had -- `credential.helper=store` rewrites by rename-over-target, which is
EBUSY on a bind mount and reports a fatal that is not one.
* Budget split deliberately: decoder 5 cpu / 6 GB, port 3 / 4, leaving room for
the planned Referee. "Half the host" was right when there was one agent.
Prompts move to docs/agents/ and are rewritten around the protocol: the oracle
is the running game, dynamic RE stays with the decoder, each iteration must
attempt to refute one claim of the other, and neither may verify its way out of
its own role.
Merges the Godot port into the reverse-engineering repository, preserving both
histories -- 1019 commits of corpus plus the port's 31, brought in by subtree
merge and then moved into place so git can follow each file across the rename.
The reason is not tidiness. The two-repo split forced the exporter to depend on
the decoders by pinned revision, and that created a whole class of failure that
now disappears: a sha reachable only from a topic branch, orphaned by a
squash-merge, breaking a fresh checkout silently at build time. It also forced a
live read-only mount of one agent's working tree into another's container, which
is why a contract file could move mid-iteration. With a path dependency, a
decoder change and the exporter change it requires land in the same commit or
not at all.
Canary stays separate: it is a fork tracking upstream.
New structure for the long term:
docs/game/ how the game is NAVIGATED -- menus, modals, prompts, alerts,
and in-game flight. Written so nobody rediscovers it. Mostly
open questions on purpose; the in-game tutorials are the
resource for the flight half.
docs/port/MODDING.md
modding as a constraint on the exporter TODAY, not a later
feature: one logical asset in one file (the disc splits nearly
everything, and resolving that is the exporter's job), names a
person recognises, PNG/OGG/OGV/JSON only, base-and-overrides so
re-exporting is always safe, provenance in every file.
data/base + data/mods
generated tree and drop-in overrides, both gitignored
exchange/ transient inter-agent files, deliberately outside history
docs/agents/ the team protocol
Both the README and the navigation doc lead with the correction that cost the
most: the oracle is the real game under Xenia Canary. Reborn's renderer is a
hypothesis under test, it has been wrong, and treating it as ground truth
propagated into three documents and both agents before a human caught it.
Scripted modding stays possible without being built: no screen name is hardcoded
in GDScript and there is no native code in port/, which is what Godot Mod Loader
needs to be able to substitute behaviour later.
