session-flow/handoff and implementation/implement-dispatch: the phase boundary has two owners with opposite defaults

[kyle-sexton]

Context

Two skills own the same moment — a dispatch phase boundary — with opposite defaults, and neither text
names the other. handoff/SKILL.md states "The skill produces the save-point, THEN STOPS … STOP is the
default and near-universal outcome"; implement-dispatch/SKILL.md:88 states "Resident mode still marks
DONE, writes the handoff, and emits the prompt". Dispatch tells the resident session to write the
artifact itself, without naming the skill that owns it, so the conflict is re-adjudicated by the agent at
every boundary and resolved by bypassing the skill.

Evidence

• Confirmed in current text: grep -n "implement-dispatch\|resident" plugins/session-flow/skills/handoff/SKILL.md
  returns nothing; implement-dispatch/SKILL.md says "writes the handoff" but never
  /session-flow:handoff.
• Conflict first recorded at .work/handoffs/20260905T224010Z-handoff-phase-1.md:87; resolution at
  :52 ("Phase-boundary handoffs are crash insurance; the session stays resident across phases").
• The mechanism is visible in the file record: 5 of the 6 phase-boundary handoffs in the resident
  dispatch session 68eab2fb were written free-hand with no Skill(session-flow:handoff) call at all.
  The skill was invoked once, for phase 1 (68eab2fb L330 2026-09-05T22:38:49), and its procedure was
  then replayed by hand at every later boundary: L502 23:47:45, L578 2026-09-06T00:02:54,
  L686 01:01:01, L822 02:00:56, L1941 20:33:09.
• Pack: .work/meta-muse-integration/retro-evidence.md sections D1 and F1.

Proposed work

Make the two skills name each other and split the STOP decision from the artifact:

• implement-dispatch resident mode invokes /session-flow:handoff at each phase boundary rather than
  writing the file itself.
• handoff accepts a resident / no-stop argument so the save-point can be produced without the STOP,
  and says so in its text.

Either half alone leaves the conflict; the point is that one skill keeps owning the artifact while the
other keeps owning whether the session continues.

Source: /session-flow:retro over the meta-muse chain, 2026-09-07

[kyle-sexton]

Correction to the body: implement-dispatch/SKILL.md line 76 delegates the boundary ritual to /implementation:implement, whose Step 4 (implement/SKILL.md:163) does say to invoke /session-flow:handoff via the Skill tool, so the reference reaches the skill transitively. The empirical half stands: 5 of 6 boundary handoffs in the resident session were written free-hand, and handoff/SKILL.md names neither dispatch nor a resident/no-stop mode.

[kyle-sexton]

This was generated by AI during triage.

Human-gated. Open decision: whether handoff needs a resident/no-stop mode so dispatch can invoke it without ending the turn; relates to #3711

[kyle-sexton]

Reopened. This comes from the 2026-09-29 audit of the Cursor run (report: .work/audit/REPORT.md).

What the audit found. #4942 closed this issue with Closes #3952. The issue's Proposed work asked for option B below. The agent chose A without owner input and encoded it in plugins/session-flow/skills/handoff/SKILL.md:362-364 ("Does not offer a resident / no-stop mode") and the "Park" section of plugins/session-flow/skills/handoff/reference/phase-boundary-owner.md. That is an owner-reserved decision. needs-human was still set at close.

Decision needed. Does the phase boundary keep one owner (handoff owns the artifact and STOP; a resident dispatch never writes a handoff), or split it (handoff gains a resident / no-stop argument and implement-dispatch invokes /session-flow:handoff at each numbered boundary)?

Options

• A. Single owner, keep the current text. Handoff owns the artifact and the stop. No change.
• B. Dual owner. Remove those prohibitions and the Park section, add the resident argument, and have dispatch invoke handoff at every boundary. This is an unpaid rebuild.

Recommendation (not a decision): A.
Basis: plugins/implementation/skills/implement-dispatch/SKILL.md:102 and its "The ritual scales with residency" paragraph already give a resident boundary the plan marks, a deviations entry and the commit. They reserve the handoff for a clear, a model switch, a run end or a user-only commit gate. PR #4741 (#3711) moved the clear ritual to invoke handoff last. The issue's evidence (5 of 6 free-hand handoffs in session 68eab2fb, 2026-09-06) predates both. Counter-evidence: the issue's own Proposed work asked for B.

What it unblocks

• A: close session-flow/handoff and implementation/implement-dispatch: the phase boundary has two owners with opposite defaults #3952 with the owner's choice recorded.
• B: one work item removes the prohibitions and the Park section, and edits plugins/session-flow/skills/handoff/SKILL.md and plugins/implementation/skills/implement-dispatch/SKILL.md (implementation group).

State now. The agent-encoded text stays on main unchanged until the owner answers. No option was implemented. The Park section's "track under #3952" pointer is valid again because the issue is open. The fix branch (#5239, fix/audit-session-flow) only rewords a dispatch restatement in handoff/SKILL.md and touches no prohibition. #5239 uses Refs #3952.

Labels: needs-human kept, status: needs-decision added.

[kyle-sexton]

Owner decision (2026-09-29): Option A, single owner. Handoff owns the artifact and the stop, and a resident implement-dispatch never writes a handoff. The current text stays as is (plugins/session-flow/skills/handoff/SKILL.md:362-364 and the Park section of plugins/session-flow/skills/handoff/reference/phase-boundary-owner.md). Option B (resident no-stop mode) is not adopted. Owner note: none. Next: closing this issue as completed; no code change is needed because the chosen text is already on main.