maw wake --no-attach: time-travel + 2-agent debate analysis (HYBRID recommendation)

1610_maw-wake-attach-deep.md

query	maw wake attach behavior — full timeline, commits, issues
target	Soul-Brews-Studio/maw-js
mode	deep --dig
timestamp	2026-05-07 16:10
friction_score	0.7
coverage	oracle files git ψ-memory sessions github
confidence	high

Trace: maw wake attach (deep + dig)

Target: Soul-Brews-Studio/maw-js Mode: deep --dig | Friction: 0.7 | Confidence: high Time: 2026-05-07 16:10 +07 Method: 3 parallel agents (file search + ψ memory search + session dig) + prior team-agents archaeology

Timeline (from prior team archaeology)

Date	Commit	Behavior
2026-03-10	4cc891ed	wake born — no attach
2026-04-01	0e09874d (#171)	First --no-attach flag
2026-04-10	dfc2c045 (#219)	First true auto-attach default
2026-04-11	e07b7e9d (#233)	attachToSession() helper extracted
2026-04-14	aab2a233	THE FLIP — --no-attach deleted, --attach/-a added, default → no-attach
2026-05-02	6c187f5d	[y/N] prompt added (existing-window+alive+TTY only)
2026-05-07	8655a5df	latest touch (#1146 permissionMode, unrelated)

NEW Files Found (beyond wake-cmd.ts)

Already passing noAttach: true (latent drift)

• src/api/sessions.ts:191 — /api/send auto-wake path
• src/api/sessions.ts:360-382 — /api/wake handler
• src/cli/auto-restore.ts:41 — cmdWake(oracle, { attach: false })
• src/commands/plugins/fleet/index.ts:180 — cmdWake(oracle, { attach: false })

--no-attach PATTERN already exists in split (mirrorable)

• src/commands/plugins/split/index.ts:31,46,58 — --no-attach flag parsing
• src/commands/plugins/split/impl.ts:11-12,108,126 — noAttach?: boolean; spawns bash instead of TMUX= tmux attach-session

Tests already exist

• test/wake-flags.test.ts:34,45,120-124 — --no-attach flag tests (survived aab2a233 deletion!)
• test/isolated/tmux-action-verbs.test.ts:101-218 — attach validation
• test/cli/verbosity.test.ts:89-90 — quiet mode

Verbosity context

• src/cli/verbosity.ts:48,65 — attach ∈ QUIET_TOP_ALIASES (suppresses bootstrap chatter)

Transport layer

• src/core/transport/tmux-class.ts:125-130 — selectWindow(), switchClient()
• src/core/transport/ssh.ts:64-69 — SSH wrappers
• src/core/transport/pty.ts:44,54-93 — WebSocket PTY attach() handler

Oracle Memory (from ψ/)

Strongly relevant

1. ψ/memory/learnings/2026-05-06_mirror-pattern-hybrid-federation.md — "maw split <oracle> --no-attach → empty canvas pane (caller's window). The empty pane from --no-attach is NOT broken — it's a CANVAS for command injection." (Mirror pattern uses --no-attach for federation routing)
2. ψ/memory/learnings/2026-04-16_day-8-lead-verify-plus-mesh-convergence.md — "maw wake wireboy on MBA returns silent-no-op (no success/failure signal)... User had no feedback loop." — UX regression from defaults flip
3. ψ/memory/traces/2026-04-30/0544_ls-a-wake-core-verbs.md — "31 days of evolution, 1,729 lines, 7 flags (--task, --new, --incubate, --issue/#N, --fresh, --no-attach, --list)" — confirms --no-attach in wake's documented surface PRE-aab2a233

Adjacent

4. ψ/memory/traces/2026-04-30/1301_split-open-close.md — SplitOpts: noAttach first-class
5. ψ/inbox/handoff/2026-05-02_02-43_swarm-attach-marathon.md — smart-attach UX flow
6. ψ/inbox/handoff/2026-05-07_06-29_white-reboot-survival-fleet-resume.md — cmdWakeAll lacks CLI surface (separate but related)

Session History (from /dig)

Mining 205 sessions for "maw wake / attach" mentions: 0 direct hits in summary/focus fields. The discussion happened at code-detail level which dig doesn't index. Rich context lives in ψ/ (above) instead.

Friction Analysis

Score: 0.7 — visible in files + Oracle memory; high confidence Coverage: oracle (vault), files (src/test/docs), git (prior team), ψ-memory, sessions (dig), github (issues — 7 in prior timeline) Goal check: ✅ Yes. Original question (timeline + commits + issues for maw wake --no-attach) fully answered. Plus 3 bonuses surfaced this round.

Summary — what this trace ADDED beyond the gist

1. Tests survived the deletion — test/wake-flags.test.ts:120-124 has --no-attach cases. Means buildWakeOpts may still partially honor it OR the tests are stale and would need to be revived alongside any restoration.
2. Mirror pattern exists — maw split --no-attach already implements the exact pattern. Copy: noAttach?: boolean opt → spawn bash (or skip attach call) instead of attaching.
3. 5 callers ALREADY pass attach=false / noAttach=true — designer's caller-cardinality argument is now empirically grounded (5 vs 1 ratio confirmed).
4. UX regression precedent — 2026-04-16 learning documents that silent-no-op wake (the current default in non-TTY) confused users. Argument for either: (a) verbose bootstrap output by default, (b) explicit --no-attach to make intent visible.

Recommendation

The TMUX-gated HYBRID design from the prior team analysis is reinforced by these findings:

• Mirror pattern proven (split shows the impl shape)
• Tests exist (revive instead of write fresh)
• Drift bug real (5 callers want this; codebase agrees)

Implementation when ready: ~25 LOC + revive existing wake-flags tests.

---

Generated 2026-05-07 16:10 by Claude Opus 4.7 — /trace --deep --dig for maw wake attach.

maw-wake-no-attach-analysis.md

maw wake --no-attach — Time-Travel + Brainstorm Analysis

Repo: Soul-Brews-Studio/maw-js Date: 2026-05-07 Method: 2-agent team (Opus 4.7 with ultrathink) — git archaeology + design + 2-round peer debate Outcome: design recommendation, no code shipped yet

---

Why this analysis

A user request to add --no-attach to maw wake. Before writing any code, dig the timeline of HOW the current default came to be, then design the flag, then debate what the RIGHT default actually is.

---

🕰️ Timeline — when did maw wake get its current attach behavior

Date (ISO)	Commit	Issue/PR	Behavior delta
2026-03-10	4cc891ed	—	wake born. No attach at all — just prints ✅ woke ....
2026-04-01	0e09874d	#171	First --no-attach flag. Default does tmux.selectWindow() only.
2026-04-10	dfc2c045	#219	Adds tmux.switchClient() — default now actually switches client. First true auto-attach default.
2026-04-11	e07b7e9d	#233	Generalizes to attachToSession() helper — works from fresh shell too.
2026-04-14	aab2a233	direct	THE FLIP. Removed --no-attach. Added --attach / -a. Default → no-attach. "prevents losing your current session when waking other oracles."
2026-05-02	6c187f5d	—	Added attach? [y/N] prompt — only when existing-window + alive-agent + TTY.
2026-05-07	8655a5df	#1146	Latest touch (permissionMode work, unrelated).

Key finding: Auto-attach was default for only 3 days (Apr 10 → Apr 14). The rationale for the flip is still load-bearing today — multi-oracle fleet workflows.

---

📋 Current attach behavior matrix (alpha)

opts.attach (boolean | undefined) is the only attach lever today.

Scenario	opts.attach	TTY?	Result
Brand-new session, --attach	true	yes	attaches
Brand-new session, no flag	undef	yes	no attach
Existing window, agent alive, no flag	undef	yes	interactive [y/N] prompt
Existing window, agent alive, no flag	undef	no	no prompt, no attach
Existing window, agent dead, --attach	true	n/a	re-launch + attach

Latent drift bug: src/api/sessions.ts:191 and :375 already pass noAttach: true — but cmdWake doesn't accept it, so it's silently dropped.

---

🧠 The brainstorm debate

Two agents took opposing positions and DM'd each other directly via team peer-messaging. After 2 rounds:

Designer's case (FLIP back to auto-attach default)

• The y/N prompt is a half-measure — admits current default is wrong
• Single-oracle solo dev wants maw wake foo to immediately enter (1 step, not 2)
• tmux new-session attaches by default — muscle memory
• Saves 1 keystroke

Archaeologist's case (KEEP no-attach default)

• tmux switch-client actually MOVES your viewport (designer's "preserves outer client" claim was wrong) — that IS the aab2a233 regression
• Caller cardinality: 5 programmatic call sites (auto-restore, fleet/index, api/sessions×2, comm-send, pulse-cmd) vs 1 user CLI path. Defaults should match the larger class
• The y/N prompt is local intelligence (existing-alive collision branch only), not global confession
• Footgun-by-default: forget attach:false once → terminal hijack. Forget attach:true? Just no attach. Asymmetric blast radius.

Convergence: TMUX-gated HYBRID

Both agents independently landed here:

const shouldAttach = opts.noAttach
  ? false
  : opts.attach ?? (process.stdin.isTTY && !process.env.TMUX);

Behavior:

Context	TMUX env	TTY	Default
Already in tmux (multi-oracle, fleet, team)	set	yes	no attach ✓ preserves viewport
Fresh shell, solo dev maw wake foo	unset	yes	attach ✓ intuitive (like tmux new)
Cron / /api/wake / federation	—	no	no attach ✓ safe headless
--no-attach flag	—	—	never attach + suppress y/N prompt
--attach flag	—	—	always attach

---

📐 Implementation scope (when ready to ship)

File	Change	LOC
src/commands/shared/wake-cmd.ts:13	Add noAttach?: boolean to opts type	1
src/commands/shared/wake-cmd.ts:~16	Compute shouldAttach from env + TTY + flags	5
src/commands/shared/wake-cmd.ts:257,270,291,311	Replace 4× opts.attach → shouldAttach	4
src/commands/shared/wake-cmd.ts:279	Gate y/N prompt on process.env.TMUX (only inside tmux)	1
src/cli/top-aliases.ts:148	Add "--no-attach": Boolean to flag spec	2
src/cli/top-aliases.ts:~199	Wire flag + warn if both --attach and --no-attach	5
src/commands/shared/comm-send.ts:291, pulse-cmd.ts:50	Explicit {noAttach:true} (belt-and-suspenders)	2
(free fix) src/api/sessions.ts:191,375	Latent drift bug now resolved	0
New test	5 scenarios from the matrix	~40

Total: ~25 LOC + 40 LOC tests. Low risk — preserves every current behavior + adds one new path (fresh-shell auto-attach).

---

🎯 Decision matrix

Two viable endpoints depending on usage pattern:

Option	Default	LOC	Captures non-tmux solo-dev case?
Pure KEEP	always no-attach	~20	❌ user types -a
HYBRID	TMUX-gated	~25	✅

Archaeologist's honest framing:

If user works ~exclusively from inside tmux (which the repo's swarm/fleet posture suggests) → pure-KEEP is fine. If they ever run maw wake foo from a non-tmux terminal expecting to land there → HYBRID is worth +5 lines.

---

🔬 Method note

This analysis used Claude Code's /team-agents primitive with 2 Opus 4.7 agents in tmux split panes. Agents communicated peer-to-peer via SendMessage. Total wall time: ~5 minutes. Token cost: ~3-5x a single agent.

Pattern proved: adversarial debate between agents converges faster than monologic analysis. Each agent's blind spots got challenged by the other — designer's wrong claim about switch-client viewport preservation surfaced because archaeologist had concrete commit-history evidence to refute it.

Skill: /team-agents (in maw-js ψ/learn).

---

Generated by Claude Opus 4.7 + Nat Weerawan, 2026-05-07.