Happier: experimental Claude pilot

Table of Contents

Result and scope

The disposable Savio pilot supports an opt-in Claude adapter using persisted history and remote startup. Both direct and persisted modes retained context through API, terminal, and browser turns. Direct mode introduced duplicate browser rows after a local-mode handoff; persisted mode displayed each measured turn once. Native transcripts and compact CLI histories contained one copy in both modes.

Phase Direct browser copies Persisted browser copies Native turn copies
Initial remote READY, including reported reload 1 1 1
Terminal LOCAL after switching to local 2 1 1
Browser BROWSER after returning to remote 2 1 1

This is an experiment adapter, not a built-in Happier harness. SuCoder's launcher, context injection, transcript probing, and tmux ownership stay in use. The background is the research note and issue 58.

Gate Measured result
Released CLI on Savio glibc 2.28 Passed with CLI 0.2.17 and Claude 2.1.310
Context and API/terminal/browser continuity Passed, with distinct remembered nonces
SuCoder framed message waits Three correlated replies, including a browser nonce
Standard sucoder attach Passed in the first pilot
Separate happier attach Passed with explicit private/default tmux metadata; failed without it
Native requested UUID SDK chose another UUID; existing cwd/frame fallback recovered replies
Default local startup Failed: missing ensureMetadataSnapshot; remote startup works
Manual browser tool approval A pending request and one successful execution were observed; decision attribution is incomplete
Allocation release Pilot processes independently observed absent
Named tmux sockets, concurrent nodes, renewal, Codex Subsequent gates

Pinned installation

The tested CLI is cli-v0.2.17, commit a2c45c86804b24e60ac0b76146a3ef36dfcaa15b, published 2026-10-22. The 0.2.19 releases checked during the pilot were UI bundles. Do not substitute a UI version for the CLI pin.

Linux asset Pinned SHA256
happier-v0.2.17-linux-x64.tar.gz 969328ed5296119d18e3ca48839e0baafb5d05d1ed7047afacfbac064fd8cbc4
happier-v0.2.17-linux-arm64.tar.gz 80e3110dc55a2a22f0cca6d26903715b67581eb5509be4e4a403b5898214bb44

install accepts an absolute, fresh destination and an optional cached release tarball. Otherwise it downloads the exact pinned asset. It checks SHA256 before unpacking, changes no shell profile, installs no service, and does not update Claude. Its version probe uses a temporary home. The checksum is pinned from the upstream HTTPS release; the helper does not verify the separate minisign signature.

Cache the compressed archive on shared storage and unpack it on local disk. There are about 31,000 archive entries. Keep the full runtime payload for agent use: the executable needs adjacent launchers and dependencies.

Adapter contract

SuCoder detects the provider from the executable basename. The isolated claude adapter preserves the native Claude launch profile and forwards model, system prompt, and requested --session-id arguments. The measured SDK UUID drift means forwarding does not guarantee runtime UUID identity.

HAPPIER_CLAUDE_PATH must name the real Claude executable by absolute path; a recursive adapter path is refused. Standalone --version and --help probes go to native Claude without an allocation, pairing, or daemon.

control requires a numeric SLURM_JOB_ID and an existing writable, absolute SUCODER_LOCAL_ROOT. It supplies:

Setting Pilot default
HAPPIER_HOME_DIR $SUCODER_LOCAL_ROOT/happier-pilot-$SLURM_JOB_ID, mode 0700
HAPPIER_NO_BROWSER_OPEN 1
HAPPIER_CLI_UPDATE_CHECK 0
DISABLE_AUTOUPDATER 1, for native Claude
HAPPIER_TRANSCRIPT_STORAGE persisted
HAPPIER_SESSION_AUTOSTART_DAEMON 0
control claude startup --happy-starting-mode remote --permission-mode default

Explicit provider options follow the pilot defaults, allowing deliberate reproductions with --happy-starting-mode local. An explicit HAPPIER_TRANSCRIPT_STORAGE=direct enables the echo comparison. Values other than direct and persisted are rejected before starting Happier. An inherited conflicting home or a symlinked pilot home is refused.

SLURM_JOB_ID is a launch guard, not proof of confinement. Record the runner's and daemon's actual /proc/self/cgroup during acceptance. Both entry points inherit CLAUDE_CONFIG_DIR and provider authentication; no credentials are copied into the repository or evidence bundle.

Set SUCODER_HAPPIER_PILOT_ARCHIVE to the cached release tarball. On first use the control helper installs it into $SUCODER_LOCAL_ROOT/happier-cli-0.2.17. Alternatively, SUCODER_HAPPIER_PILOT_CLI may name an already unpacked absolute executable; the caller is responsible for that executable's provenance.

A daemon is optional for a terminal-started session. When needed, use control daemon start-sync in another allocation-owned tmux window. The helper rejects setup/service installation, detached daemon start, and daemon stop --all. A shell, tmux, or setsid does not escape Slurm's allocation cgroup.

Run another bounded pilot

Create a throwaway canonical repository and private config based on the existing cluster target. Keep normal project mirrors out of that config. Preserve its account, partition, QOS, and certificate; use a short time limit, confined: true, local_disk: true, wip_snapshot_minutes: 0, and codex_state: {enabled: false}. Upload the complete scripts/happier-pilot/ directory and compressed release to a private remote pilot directory.

agent_launcher:
  command: [/absolute/private/pilot/adapter/claude]
  model: sonnet
  needs_yolo: false
  env:
    HAPPIER_CLAUDE_PATH: /absolute/path/to/the/real/claude
    SUCODER_HAPPIER_PILOT_ARCHIVE: /absolute/private/pilot/happier-v0.2.17-linux-x64.tar.gz
    HAPPIER_TRANSCRIPT_STORAGE: persisted

The mirror prompt should contain a unique context marker and SuCoder's REPLY <id>: convention. Use no tools during continuity turns. Keep the native transcript for independent message and tool-execution counts.

python -m pytest -q tests/test_happier_pilot.py tests/test_reply_launch.py
sucoder --no-agent-sudo -c "$PILOT_CONFIG" -T carleton-htc agents-clone "$PILOT_MIRROR"
sucoder --no-agent-sudo -c "$PILOT_CONFIG" -T carleton-htc --time 00:45:00 agents-run "$PILOT_MIRROR" --no-sync
sucoder --no-agent-sudo -c "$PILOT_CONFIG" -T carleton-htc attach "$PILOT_MIRROR"

Here --no-agent-sudo uses the human account for the laptop-side access check; remote execution still uses the target's SSH account. The library path is ensure_remote_clone(context) followed by launch_agent(context, sync=False, detached=True); local ensure_clone does not stage a remote mirror.

Complete the original terminal's pairing request in the human's browser. Do not start parallel authentication requests for the same home. Fresh per-allocation homes require fresh pairing; do not archive their access keys. Record the job id, node, private tmux identity, Happier id, requested Claude UUID, and actual provider transcript UUID separately.

Use the same control helper, inside that allocation, for:

/absolute/private/pilot/adapter/control session list --json
/absolute/private/pilot/adapter/control session history "$HAPPIER_SESSION" --json
/absolute/private/pilot/adapter/control session status "$HAPPIER_SESSION" --live --json
/absolute/private/pilot/adapter/control session send "$HAPPIER_SESSION" 'Report the context marker and remembered nonce without tools.' --wait --timeout 90 --json

Verify SuCoder's own framed reply path before and after browser input:

sucoder --no-agent-sudo -c "$PILOT_CONFIG" -T carleton-htc message "$PILOT_MIRROR" \
  'Report the marker and nonce. Answer with REPLY <the incoming message id>: ...' --wait 120

For a display comparison, use distinct direct/persisted working directories and nonces. Observe initial remote turns, a terminal LOCAL turn after Ctrl-T, and a browser BROWSER turn after switching back to remote. Record browser counts before and after a full reload; count native frames separately. Persisted mode stores encrypted relay history; native Claude JSONL is still written and remains available to SuCoder's transcript probe.

Attachment boundary

Daemon-first alone did not make the unregistered control attachable. Separate Happier attachment passed when the session was started with:

--happy-terminal-mode tmux --happy-terminal-requested tmux \
--happy-tmux-target SESSION:WINDOW --happy-tmux-tmpdir PRIVATE_DIRECTORY

Those checks used a dedicated default socket under $TMUX_TMPDIR/tmux-UID/default. Happier's attach plan assumes that socket name. SuCoder uses a named socket, so this does not establish named-socket compatibility. Standard sucoder attach passed separately in the first pilot and remains the measured route for the SuCoder-owned terminal.

Approval evidence and remaining gate

The final persisted test requested exactly one harmless Bash call:

python3 -c 'print("HAPPIER_APPROVAL_R3_OK")'

Live status first showed one pending request. The native transcript had a tool call without a result. A later result printed the marker once, and pending requests returned to zero. The human did not see an approval request. A pending count is not evidence that a browser card was visible.

Retrospective inspection of the complete native JSONL found:

  • A hook_permission_decision attachment at 09:54:33 Pacific, with PermissionRequest, decision: allow, and the matching tool-use id.
  • The successful result at 09:54:41 records permissionDecision: {decision: accept, source: hook, reasonType: hook, reasonCode: bashMiss:no-rule-match}.
  • Native permission-mode records say auto, while Happier's status said default. Status alone is not authoritative for every permission layer.

This identifies a hook grant. A subsequent read-only audit, authenticated through a temporary client, recovered the encrypted completed-request record:

Field Recorded value (2026-10-24, Pacific)
Request created 09:53:10.145
Marked push-notified 09:53:10.148
Request completed 09:54:33.230, about 83 seconds later
Status approved
Pending request absent
Completed permission requests in this session 1
Permission mode change, saved allow rules, decision reason absent from the completion

Happier's metadata still reports the original default permission mode and its original timestamp. The suggested allow rule in the request is not a record of a rule being saved. The record's shape is consistent with a one-time permission response. In the pinned bridge, expiry is recorded as cancellation, and mode-based auto-completion includes its mode. Thus the record does not support a timeout-to-allow explanation.

The approving device/actor is not retained in this record, so it cannot establish which client or action supplied the approval. Manual browser UI approval remains unverified. The temporary audit credentials were removed. Do not describe the marker as an approval-UI pass or ask the human to approve this already-settled request.

For a future synthetic approval test, retain a private diagnostic trace before starting the session:

export DEBUG=1 HAPPIER_STACK_TOOL_TRACE=1
export HAPPIER_STACK_TOOL_TRACE_FILE="$PRIVATE_EVIDENCE/tool-trace.jsonl"

Capture request id, effective permission modes, hook source, completed request status/reason, and the response time before release. Preserve the native permissionDecision and hook_permission_decision fields when sanitizing evidence. Test deny and allow separately. Do not change to YOLO or install broad allow rules to make this gate pass. Review diagnostic logs before sharing them; keep credentials and unrelated content private.

Release and evidence

Save synthetic turn counts, versions/digests, cgroups, session identities, and targeted approval records. Stop only the pilot's sessions and daemon, then release its exact allocation:

sucoder --no-agent-sudo -c "$PILOT_CONFIG" -T carleton-htc release "$PILOT_MIRROR" --force

Check scheduler termination and, from a separate authorized allocation on the same node, absence of the pilot's process identities. Check the browser offline indicator separately; scheduler state does not prove that UI state. Retain the compressed archive and sanitized evidence, not account keys or the disposable runtime home.

The first pilot job 40423899 was released; observer 40425532 saw its processes disappear. Follow-up job 40429332 timed out. The successful comparison ran as 40439415 on n0035.savio4 and was cancelled at 10:00:17 Pacific on 2026-10-24. Observer 40439597 saw live processes, then two empty scans, and completed successfully at 10:00:21. All recorded Happier pilot allocations from the previous day were also confirmed ended. Other mirror allocations were not changed.

Private local and Savio evidence is under ~/.sucoder/pilots/happier-ab-20261024T163620Z/. It includes browser-observations.json, native-final.json, compact histories, approval-decision-audit.json, completed-approval-audit.json, observer-40439597.jsonl, and final allocation accounting. Earlier SuCoder message evidence is under ~/.sucoder/pilots/happier-20261024T000415Z/.

References

SuCoder — Home · GitHub