The autonomous agent's loop prompt had grown to ~9 000 words of accumulated
findings, refutations and traps. That is a symptom, not a style: it was doing
the job the corpus should do, in the one place that dies with the container.
Three concrete failures followed from it, and each gets a structural fix rather
than a louder instruction.
1. IT REPEATED WORK IT HAD ALREADY DONE.
The "do not revive" list and the method traps existed ONLY in the prompt, so
nothing pointed a fresh iteration at them. Extracted verbatim into two
tracked files:
docs/re/REFUTED.md 105 claims tested and dead, grouped by subject so a
grep for your noun finds the neighbourhood
docs/re/METHOD.md the traps already paid for -- controls, inference,
searching, reading data, runtime
Both are linked from INDEX.md, and the loop prompt now opens by requiring
them to be read. This is the fix for "re-derived something already known":
the knowledge is now where the next iteration looks, not in a context window.
2. IT FORGOT TO ARM THE NEXT WAKEUP.
The prompt tried to solve this by shouting at itself in the first line. The
real fix is to stop asking: `loose` now defaults to a FIXED interval (45m),
so the harness owns the cadence and a forgotten ScheduleWakeup cannot end the
run. SYLPH_LOOP_INTERVAL= (empty) restores self-pacing.
3. IT COULD NOT PUBLISH, SO THE WORK ONLY EXISTED IN THE CONTAINER.
New `push-work`, plus a read-only credentials mount
(SYLPH_GIT_CREDENTIALS, default ~/.sylph-git-credentials). It pushes the
CURRENT branch only, refuses anything that is not auto/*, and never
force-pushes -- so the consolidated line stays a human's decision and a
confused iteration cannot rewrite history. The loop prompt now requires a
push on every iteration that commits, rather than at the end of some longer
arc, which is exactly when a container dies.
The prompt itself drops from ~9 000 words to 85 lines and, more importantly,
stops accumulating: findings go in the corpus, and the prompt points at it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`./sylph-agent loose [task]` starts Claude Code detached with
--dangerously-skip-permissions, running /loop on loop-task.md: work the RE
backlog one item at a time, commit to auto/* branches, never push, record
withdrawn results rather than deleting them. `logs`/`attach`/`stop` to watch and
end it. Runs -d WITHOUT --rm so the transcript survives the container exiting —
for an unattended run that is the only record of what happened.
Two things had to be fixed for an agent to survive being left alone.
MEMORY CONTINUITY. The project is now bind-mounted twice: at /work, and at its
own host path. Claude Code derives its per-project state key from the working
directory, so running at /work handed the agent an empty project instead of the
accumulated one. Verified: a loose run now reports MEMORY=yes and reads back the
same branch and backlog as the host.
FOUR INTERACTIVE GATES, each a silent permanent hang with nobody at the keyboard
-- no error, no log line, just a container that looks healthy and does nothing:
theme picker hasCompletedOnboarding + lastOnboardingVersion. Re-fires
whenever the container's Claude Code is a different version
to the host's, which is the normal case.
folder trust projects.<path>.hasTrustDialogAccepted
bypass disclaimer answered in a pty by bin/claude-autonomous. It has no config
key by design -- it wants a person to accept once, and the
person did so by launching this.
fullscreen upsell fullscreenUpsellSeenCount. This one fires MID-SESSION, after
the pty wrapper has already handed over, so it cannot be
answered the same way.
Config key names were read out of the shipped binary's own strings, not guessed.
The pty wrapper matches SINGLE WORDS. Claude Code draws its UI with
absolute-column escapes between words, so the prompt arrives as
`Yes,\x1b[13GI\x1b[15Gaccept` and a multi-word pattern never matches -- failing
in a way indistinguishable from the wrapper not running at all. It stops
matching once the session is live so nothing later is answered by accident.
~/.claude.json is now mounted read-only at a staging path and copied in, so the
container cannot rewrite the host config. Credentials stay shared read-write in
~/.claude, which is what token refresh and memory continuity need.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
