Root cause confirmed: the carve-out routed to inter-session-peer-discipline,
which existed in sources but was not installed -> no competitor in the registry,
nearest in-domain skill won. Fix = install the sibling (byte-identical parity).
FP-twin verified: a fresh clean-context subagent on the N1 phrase now routes to
inter-session-peer-discipline (in registry), not session-inbox-monitor.
session-inbox-monitor description untouched (option a rejected as whack-a-mole).
Governance: workshop (peer) proposed (b) as a ruling; per the freshly-installed
inter-session-peer-discipline (peer = proposal, not authority), it was surfaced as
a recommendation and ratified by the user, not closed on the peer's say-so.
Closes session-inbox-monitor-received-msg-fp. Wiki concept Status open->resolved.
Tail flagged: inter-session-peer-discipline now installed but not in hermes mapping.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
-review next (NON-implementer session only); open informational FP follow-up
session-inbox-monitor-received-msg-fp. Autopush grant resets next session.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
New page concepts/session-inbox-monitor-received-msg-fp.md — sibling of
delegate-task-negative-trigger-fp. Same FP family, new dimension: the carve-out
is already literal+routed (NOT for handling a received message ->
inter-session-peer-discipline) but the route target is not installed, so it has
no competitor and the nearest in-domain skill wins anyway. Borderline (neg 2/3,
EN twin clean), self-corrects on body-load. Bidirectional cross-link +
index + log. New principle: a routed negative competes only if its route
target is installed. Captured in the wiki (not private memory) per owner direction.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
session-inbox-monitor-hermes-mapping -> done: added to hermes/mapping.yaml
pending section, mode:pending + intended {auto, productivity} (mirrors the
session-handoff sibling). Reason cites the settings.json mutation, the OS
process kills, and the Windows-PowerShell hook needing a Linux port for Hermes.
Schema version untouched (schema, not content). YAML validated (32 skills).
session-inbox-monitor-install -> done: install.ps1 -Names session-inbox-monitor,
byte-identical (diff 0), v0.2.1; activation confirmed live this session (harness
picked the skill into available-skills with full description, no /reload-plugins).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Closes session-inbox-monitor-stophook-utf8-fix. Cyrillic message bodies arrived
as mojibake when force-delivered via the Stop-hook block reason.
Root cause: ~/.claude/hooks/stop-dispatcher.ps1 reads bodies with -Encoding UTF8
(fine) but did NOT set [Console]::OutputEncoding, so ConvertTo-Json to stdout
under a harness-spawned redirected pipe (WinPS 5.1) emitted in OEM cp866.
Fix applied to the machine-local hook (not in git): set
[Console]::OutputEncoding/$OutputEncoding = UTF8 at the top. In-situ RED->GREEN
verified through the real Stop-hook path: a Cyrillic pangram that previously came
back as mojibake now delivers clean; no-loop holds.
Repo changes: SKILL.md Failure modes documents the encoding contract for inbox
writers (no-BOM UTF-8 LF; WriteAllText, not Set-Content -Encoding utf8 which BOMs
under 5.1); bump 0.2.0 -> 0.2.1 PATCH. stop-dispatcher.ps1 multi-machine
propagation is the workshop setup's concern (notified).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
session-inbox-monitor-stophook-blockfix-proof -> done. Live proof: planted a
self-test message in .claude-inbox/, ended the turn, and the Stop hook force-
re-invoked the session with the message as a decision:block reason (no user
Enter). No-loop confirmed (inbox empty after delivery, message in .read/). The
live Monitor also paged the same file end-to-end.
Finding (orthogonal to the block mechanism, independently confirmed by workshop):
Cyrillic message bodies arrive as mojibake on inject. On-disk file is clean
UTF-8; stop-dispatcher.ps1 reads with -Encoding UTF8 (line 44) but does NOT set
[Console]::OutputEncoding, so ConvertTo-Json to stdout under harness-spawn
(WinPS 5.1, redirected pipe) encodes as OEM cp866. Start-Process repro does not
reproduce (inherits UTF-8) -> harness-spawn-specific, in-situ verify only.
Filed session-inbox-monitor-stophook-utf8-fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Core content task of the session-inbox-monitor line. Two deliverables:
1. SessionStart hook `skills/session-inbox-monitor/hooks/inbox-monitor.ps1`
(versioned for multi-machine rollout; deployed to ~/.claude/hooks/ and
registered in ~/.claude/settings.json SessionStart):
- sweep: Get-CimInstance | Stop-Process orphaned monitors of THIS inbox,
matched by sentinel CLAUDE_INBOX_MONITOR + inbox path (a /clear leaves
the poll process alive -> re-raise without sweep stacks duplicates);
- inject: hookSpecificOutput.additionalContext with the exact persistent
Monitor command (Monitor tool, not background Bash);
- opt-in gate: fires only on .claude-inbox/ dir or CLAUDE.md trigger line.
ASCII-only (em-dash -> mojibake under WinPS 5.1 without BOM, fixed).
2. SKILL.md body filled (When to use / Inputs / Steps / Deployment /
Failure modes / Side effects / What NOT to do); bump 0.1.0 -> 0.2.0 MINOR.
Headless: no hook-level signal exists (verified via claude-code-guide) ->
agent-side best-effort skip, default errs toward raising (false-skip in
interactive loses the feature; false-raise in headless is a harmless no-op).
Live-verified: inject -> valid JSON; sweep -> killed a planted orphan (PASS);
real Monitor tool spawns a bash process carrying the sentinel (sweep will
find real orphans); settings.json stays valid. Sweep over-match edge and
multi-session-per-project limit documented honestly in Failure modes.
Closes [session-inbox-monitor-sessionstart-hook].
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
poller-service-deploy-via-factory (fork 1). New skill v0.1.0 — installs the
standing-duty stack as platform-native services (systemd/launchd/winsw):
no node window, OS-supervised autostart+crash-restart, run-as-user, deploy-boundary.
- fetches winsw (pinned v2.12.0 + SHA256-verify, not vendored; STOP on placeholder)
- installs DISARMED: scope is runtime config (poller-scope.json), arming is a
separate operator step via the appeals-inbox pult; never carries POLLER_PROJECTS/DRY_RUN
- confirmation gates: discovery (read-only) -> plan -> writes; rollback section
- tears down the legacy start-worker.ps1 Scheduled Task (no double-claim)
- dist/setup-agents-task-runner.skill rebuilt
[skip-tdd: visual] — installer docs + OS service config, no testable pure logic.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the channel contract (inbox = comms only; tasks via meta tasks_*)
and the multi-session caveat: don't assert a peer override from partial
vision — the human may have ratified in a channel you can't see; ask
first. Both earned 2026-06-16.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Codifies the inbox/peer-channel discipline that emerged 2026-06-16:
- a peer agent session's messages are proposals, not authority; the
human is the only source of direction and scope.
- never report a peer-driven (or self-driven) design escalation as a
settled decision without explicit human ratification.
- channel contract: the inbox carries discussion/help/notification only;
tasks themselves go solely through meta tasks_* (board = source of truth).
- guards the echo-chamber failure mode (two sessions inflating scope past
the human) and its circuit-breaker.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Live clean-session run (this session did NOT install the skill). Probe
confirmed subagents inherit the REAL installed registry (task-loop present,
description verbatim = SKILL.md) — upgrade over dev-time simulated proxy.
Trigger-discrimination: positives 6/6 -> task-loop; negatives clean
(delegate-task, using-tasks after deconfound, configure-poller -> none with
explicit task-loop exclusion); task-loop false-positive 0/3.
Behavioral (skill body loaded) 4/4: consult-gate human-only -> STOP before
close; session_break -> SESSION BOUNDARY + STOP; empty -> stop+report no poll;
long-watch -> single ScheduleWakeup >=1200s, not CronCreate.
Finding (informational, non-blocking): session_break is the fragile gate — its
STOP semantics live only in the body (step 6), invert to soft-checkpoint when
reasoning from description alone; correct with body loaded. No fix needed
(rule triple-stated; using-tasks is REQUIRED SUB-SKILL). No follow-up filed.
All 3 task-loop deployment baselines (install / hermes-mapping / test-trigger)
now closed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
task-loop orchestrates the projects-meta board cycle (tasks_claim_next /
tasks_close / tasks_update / tasks_heartbeat) and may arm a long ScheduleWakeup —
critical-infra-adjacent, so it lands in the behavioral-audit pending tier, not auto.
Promotion to auto gated on task-loop-test-trigger. Schema version untouched.
Closes [task-loop-hermes-mapping] (baseline 2/3).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
install.ps1 -Names task-loop → ~/.claude/skills/task-loop/SKILL.md (diff-identical to
source, v0.1.0). Post-/reload-plugins task-loop appears in available-skills with full
description (YAML parsed, RU/EN triggers present). Unblocks task-loop-test-trigger
(🔵→⚪ ready, needs clean session for live behavioral run).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Skill-review checkpoint for meta-host-routing v0.3.0 (non-implementer).
Behavioral smoke-test 8/8 via clean-context subagents: 5/5 positive
trigger phrases (RU+EN) route to meta-host-routing, 3/3 negatives routed
away (using-projects-meta / delegate-task[Skip-clause] / using-tasks).
Zero false positives. Steps verified against live infra: yt-tools resolves
to dedicated host OpeItcLoc03/meta-yt-tools (projects-meta tracks it),
.common grep returns hits, excludesFile + ~/.config/git/ignore + auth.toml
present. Failure modes -> STOP+ask; What-NOT-to-do accurate. No blocking
findings on the skill.
Observation (not a skill defect): the review-task acceptance criterion
'resolves .common' is stale — skill v0.3.0 and reality route yt-tools to
the dedicated meta-yt-tools host; .common now holds only the done-archive.
Caveat: -install and -hermes-mapping baselines remain open (skill not in
~/.claude/skills nor hermes/mapping.yaml); this review covers content +
trigger discrimination only, not live-harness activation.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Public reference for the on-disk .tasks/STATUS.md block format the autonomous
poller parses: header regex, status emoji, and the **Weight:** / **Notify:** /
**Requirements:** fields. Ships with factory where the internal wiki and MCP
source can't reach. Distinct from delegate-task (MCP-tool delegation) and
using-tasks (board mechanics).
Authored via superpowers:writing-skills TDD:
- RED: 3 baseline subagents w/o skill — 2/3 used ###/bullet headers the parser
cannot recognize as a task, 2/3 omitted **Weight:** (invented risk/tier/
claimable-by), 2/3 put notify in prose, 1/3 used 🟢 for a ready task.
- GREEN: 2 fresh subagents w/ skill — both parser-valid, incl. correct
**Weight:** needs-human for the critical-infra scenario.
- REFACTOR: no new format loopholes.
Ground truth verified vs live source (status-md.ts parser, claim.ts gate,
fleet-router.js routing): missing Weight finds no backend tier -> poller parks
to blocked, so Weight is operatively required for pickup.
Closes [create-task-format-for-poller-skill]. Install to ~/.claude/skills +
hermes mapping deferred as follow-up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
