V3.6 is an infrastructure release. No new lifecycle commands, no new agents — instead, the plumbing
that was missing: a context monitor that tells you when to compact before the window collapses,
a persistent audit trail so learnings survive across sessions, and a live chat dashboard so you can
watch the squad work in real time. GSD workflow integration was also removed — the Review Squad
is now a fully standalone tool.
-
WARNING at 65% context used
A PostToolUse hook fires on every tool call and checks the current context window usage. At 65%, it surfaces a WARNING in the terminal — a reminder to run
/compact Focus on [active feature] before the window fills.
-
CRITICAL at 75% context used
At 75%, the hook escalates to CRITICAL. This is the last safe window to compact with meaningful fidelity — waiting for auto-compaction at 95% risks losing verbatim content (stack traces, config values, function names) that doesn't survive the generic summary.
-
Always use
/compact Focus on [feature]
The hook message reinforces targeted compaction over bare /compact. A topic-focused prompt biases the summary toward the facts you actually need to carry forward, rather than a generic digest of everything that happened.
-
GSD workflow integration removed
The Review Squad no longer has any dependency on the GSD workflow system. All cross-references, conditional logic, and shared state between the two systems have been removed. The squad installs and runs independently.
-
Self-contained installation
A single
curl installs /update-reviewsquad, and that command handles everything — agents, commands, hooks, templates. No external system needs to be present or configured for any command to work.
-
Findings saved to
learnings.jsonl
After each /audit run, PM Cory appends structured findings to .review-squad/<project>/learnings.jsonl — an append-only log that accumulates institutional knowledge across sessions. Future reviews can reference this history rather than rediscovering the same issues.
-
Summary saved to
review-history.md
A human-readable summary of each audit (date, scope, verdict, top findings) is appended to .review-squad/<project>/review-history.md. Both files are gitignored — local to the machine, never committed.
-
PM Cory reads history on every review
At the start of each review cycle, PM Cory loads the learnings and review history for the current project. Recurring patterns are flagged explicitly; findings that were already addressed don't generate noise.
-
All 25 agents can broadcast via
csend
Every agent in the squad now has the csend command wired in. Agents broadcast phase milestones, key decisions, and progress notes to the chat server as they work. Three message levels: phase (milestone), decision (key call), conversation (progress note).
-
Live dashboard at
http://127.0.0.1:4001
The chat server runs a real-time dashboard you can open in any browser while agents are running. Messages stream in as they're sent — no refresh required. The API server runs on port 4000; the dashboard on 4001.
-
/agent-chat:on — starts the server as a background daemon
Safe to run if already running — reports status and exits cleanly rather than starting a duplicate process. Writes a PID file to /tmp/agent-chat.pid that agents check before broadcasting.
-
/agent-chat:off — stops the server with optional log export
If messages were captured during the session, prompts to copy the log from /tmp to a permanent location before shutdown. Safe to run if not running — reports status and exits cleanly. Auto-save on dirty exit preserves the log even if the server crashes.
-
Agents broadcast only when server is running
Each agent checks for the PID file before calling
csend. If the server is off, csend is a no-op — no wasted tokens, no errors, no side effects. The server is entirely optional; squad runs work identically without it.