/ponytail-review latches the session mode flag, replacing the ruleset with a pointer line for the rest of the session #736

Summary

/ponytail-review is documented and implemented everywhere else as a one-shot
review command, but in the hooks it writes review into the session mode flag.
Because review is an INDEPENDENT_MODES entry, every later consumer of
getPonytailInstructions() gets a one-line pointer instead of the actual ladder
— for the rest of the session, with no automatic way back.

Detail

hooks/ponytail-mode-tracker.js:31 maps
/ponytail-review to mode = 'review', which falls through to the
setMode(mode) branch at hooks/ponytail-mode-tracker.js:65
and persists review to $CLAUDE_CONFIG_DIR/.ponytail-active. Nothing ever
clears it.

From then on, hooks/ponytail-instructions.js:79-81
short-circuits on INDEPENDENT_MODES and returns only:

PONYTAIL MODE ACTIVE — level: review. Behavior defined by /ponytail-review skill.

Affected injection points, all of which now emit that pointer instead of the ruleset:

• hooks/ponytail-subagent.js:25 — every
  Task-spawned subagent for the rest of the session runs effectively
  ponytail-unaware, which is the exact failure Ponytail rules don't reach subagents — add SubagentStart hook #252 added this hook to fix.
• hooks/ponytail-mode-tracker.js:111 —
  Qoder, which relies on UserPromptSubmit for per-turn ruleset injection, gets
  the pointer on every prompt.
• pi-extension/index.js:209 — pi's
  before_agent_start system prompt.
• .opencode/plugins/ponytail.mjs:77 —
  OpenCode's per-turn system.transform.

ponytail-mcp is unaffected: resolveMode() in
ponytail-mcp/instructions.js filters review
out already.

The statusline also sticks: hooks/ponytail-statusline.sh:17
renders any non-full mode verbatim, so the bar reads [PONYTAIL:REVIEW]
indefinitely.

Recovery is worse than it looks. Bare /ponytail is report-only since #99
(hooks/ponytail-mode-tracker.js:49-51) —
it reads the flag and echoes level: review without rewriting it. So the most
natural corrective gesture confirms the stuck state rather than fixing it. Only
an explicit /ponytail lite|full|ultra, /ponytail off, or stop ponytail
clears it.

Why this is a bug, not the design

Every doc treats review as one-shot, not a level:

• skills/ponytail-help/SKILL.md:16-22
  lists only lite/full/ultra under Levels ("Level sticks until changed or
  session end"); ponytail-review appears under Skills at
  line 30.
• README.md:308 describes it as "Review the current diff for
  over-engineering, hands back a delete-list."
• hooks/ponytail-config.js: 'review' in VALID_MODES but not RUNTIME_MODES — config accepts it but silently falls back to 'full' #377 already established that review is not a valid default —
  hooks/ponytail-config.js:78-81 and
  writeDefaultMode both refuse it. Refusing it as a persisted default while
  letting it persist as a session mode is the same class of mistake, one layer
  down.

Related inconsistency: the Claude tracker rejects /ponytail review (the arg
falls through to getDefaultMode()), but OpenCode's command.execute.before
runs it through normalizePersistedMode, which accepts review via
normalizeConfigMode. So /ponytail review latches OpenCode into the pointer
state even though the same input is a no-op under Claude Code.

Missing test coverage

tests/hooks.test.js mentions review only at
lines 462-486, and every one of those asserts
that review is refused as a default. No test drives /ponytail-review
through ponytail-mode-tracker.js at all, so neither the flag write nor its
downstream effect on getPonytailInstructions() is covered anywhere.

Suggested fix

Treat review like /ponytail-help: let it dispatch the skill without touching
the flag file. Concretely, don't setMode() for review in the tracker —
emit the mode-active line and leave the persisted level alone. That makes
INDEPENDENT_MODES unreachable through the runtime path, at which point the
branch in ponytail-instructions.js is dead code and can go too.

Tests worth adding:

• /ponytail-review leaves an existing full/lite/ultra flag untouched.
• /ponytail-review on a fresh session does not create the flag with review.
• A subagent spawned after /ponytail-review still receives the full ladder.
• OpenCode's /ponytail review does not persist review.