Originally created by: 5dive-bot
CLI half of DIVE-2818. Tracking row: DIVE-2824 (parent DIVE-2818).
Not merely merged. Deployed. This is the point of the two-row split, not a
courtesy note.
The plugin handler is inert until a human sends a matching string, and the
deployed CLI 0.19.1 already accepts --channel-msg, so the plugin half works
alone. This half shipped first invites humans to reply with nothing listening,
which is the dead-prompt failure delivered by deploy order instead of by
allowlist. A dead prompt is strictly worse than a dead button: it is discovered
only AFTER the human typed the exact string we asked for.
Sequence agreed with olivia (verifier): plugins [#20] merges and deploys -> dev3
flags the deploy -> olivia confirms -> this merges.
On a HIGH-STAKES tier-2 gate (spend / secrets / irreversible), the DM prompt
asks the human to REPLY with the exact string DIVE-xxxx <answer> rather than
only offering a button. The button remains as fallback.
--channel-msg is the one clear on this rail the filing agent cannot forge: it
cites the id of the human's OWN message and Telegram attests the authorship. It
shipped, it verifies, and it had ZERO callers, so the weak nonce path was the
only reachable one. DIVE-2802 is the live case: a clear recorded
answered_by=human:olivia uid=1011, the filer's own uid.
The cheapest trial of the strong path was run and it lost. lodar was invited to
answer DIVE-2802 by replying, with the exact string in the ask, and he tapped
anyway. A path whose only invitation is a sentence competing with a button is
not a reachable path, which is why this makes the prompt the DEFAULT rather than
documenting the option harder.
HIGH-STAKES ONLY (DIVE-2802 gate, answered 2026-08-05). The tap stays as
fallback either way: the 1h freshness window is a hardcoded ceiling that cannot
be widened, so a mandatory strong path would reject genuine human answers
(DIVE-525: a real human tap is NEVER rejected). The reachable goal is
"sole-nonce becomes a measurable exception", not zero.
Values are strict, not coerced. Condition 5 substring-matches the human's text
against the value we pass, so accepting "approve" and helpfully sending
--value=approved buys a refusal the human reads as breakage. Every
normalisation upstream of a literal-bytes check is a lie waiting to be caught.
The secret carve-out is preserved and tightened: exactly the token provided,
never free text, and the refusal does NOT echo what was sent, because a nudge
quoting a rejected credential copies it into a second message.
tests/gate_reply_to_clear_unit.sh — 25 arms, 0.9s, core tierThe R arms are the ones worth attention. They feed the EXACT string the DM
prints to the REAL _gate_channel_session_ok, stubbing only
_gate_channel_api (the one call that leaves the box). The copy is graded by
the function that will judge it, not by a re-implementation of its rule, which
would pass forever on the day the checker tightens. Non-vacuity is pinned
beside it: a bare DIVE-N yes refuses, and a 3700s message refuses, so the
copy's specificity and its "1 hour" claim are both load-bearing rather than
decorative.
Mutation-graded, each mutant turning a NAMED arm red: removing the notify call
site, removing the scope boundary, dropping the normaliser, dropping the answer
value from the copy, and widening the allowlist back to the tap list.
No live demo. channel_session is 0 in all history, and the acceptance clause
on the PARENT (DIVE-2818) needs a real high-stakes gate cleared by lodar
REPLYING, recording evidence=channel-session. That cannot happen before both
halves deploy.
Fork parity: codex/grok/antigravity/pi/opencode have no handler and are
deliberately not prompted. Worth its own row.
_gate_filer_answered is untouched and independently shippable.
Ticket changed by: lodar