Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 28 additions & 1 deletion docs/decisions/0029-rig-reply-end-needles-are-exchange-unique.md
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,8 @@
- **Implements:** [issue #125](https://github.com/fluxvale/dashfloat/issues/125).
- **Pattern precedent:** the F7 caveat's own remedy (`wait_for` a content
needle, not quiet) — this ADR makes that remedy structurally safe.
- **Amendment:** [2026-11-02, issue #143](#amendment-2026-11-02-issue-143-the-episodes-queue-is-dropped-not-wired)
(the never-landed episode dispatch is void — the queue was deleted).

## Context

Expand Down Expand Up @@ -63,7 +65,8 @@ replay into a snapshot screen is heavy machinery for a rig).

- A future episode dispatch in the mock must pop `EPISODES` before taking a
stock number — episode replies carry their own unique text and take none
(marked at the counter in `_sse`).
(marked at the counter in `_sse`). *(Voided — 2026-11-02, #143: the
queue was dropped, never wired — see the amendment below.)*
- If the app ever retries a stock request on a new connection, the retry
takes the next number and the scenario's wait times out — a loud capture
failure, which is the rig's contract (fail, never ship).
Expand All @@ -88,3 +91,27 @@ replay into a snapshot screen is heavy machinery for a rig).
diff), no wait draws from the stock body or `REPLY_TOKENS`, the reply-end
family rides `stock_closing` (site count never shrinks), and both waits
match wrap-collapsed text. It is red on the pre-fix rig by construction.

## Amendment (2026-11-02, issue #143): the EPISODES queue is dropped, not wired

**Status:** Accepted (the owner's standing default — dead code misleads —
declared by the operating PA and not countermanded; the fork and its
reversibility are recorded in issue #143 and its closing PR, and the
morning-revert path below).
The Context's "separate follow-up" is resolved: of the wire-or-drop fork,
the rig **drops** the interaction-surface episode machinery.

- **Deleted:** `MockHandler.EPISODES`, `App.script/1`, the never-landed
vet dispatch (`VET_REPLY`, `is_vet_request` — zero callers since the
#141 era), the four unreachable scenarios (`approve-choice`,
`question-single`, `question-multi`, `auto-reject` — never in the
argparse menu), their dispatch branches, and their entries in the
needle allowlist. The sentinel-site floor (≥9) is untouched: those
scenarios waited on literals, not `stock_closing`.
- **Kept:** the shipped captures in `docs/screenshots/133–135/` as frozen
evidence — the widget behavior they prove is app behavior (ADR 0028)
with its own unit/integration tests; only the rig's ability to
regenerate them died (with the queue, not with this ADR).
- **If the owner rules wire-in-the-morning:** revert the squash — the
scenarios' text is in git history, and the wiring rule this ADR
recorded (pop before taking a number) is the design to implement.
188 changes: 2 additions & 186 deletions scripts/capture.py
Original file line number Diff line number Diff line change
Expand Up @@ -748,21 +748,9 @@ class MockHandler(BaseHTTPRequestHandler):
def log_message(self, *args):
pass # silence — the mock never talks over the capture

# The interaction-surface captures (issues #133/#134/#135): a FIFO of
# scripted episodes — ("tools", [(id, name, args_json), ...]) a
# tool-calling episode, ("text", "...") that prose. DEAD, kept only so
# the episode-dependent scenarios still enqueue: nothing pops this
# queue (the consumer was lost in the #133–#135 era; #142 deleted its
# unused corpse and this note replaced the false claim), so those
# scenarios cannot regenerate — see issue #143 for the wire-or-drop
# decision. The vet's own fresh-context completion is detected by its
# instruction and answered with VET_REPLY (issue #135's still).
EPISODES: list = []
VET_REPLY = '{"decision":"confirm","reason":"deletes files outside the project"}'
# Stock replies are numbered per App run (App.__init__ resets this):
# the n-th streamed stock reply closes with stock_closing(n) (#125,
# ADR 0029). Episode replies carry their own unique text and take no
# number.
# ADR 0029).
STOCK_REPLIES = itertools.count(1)

def _json(self, payload, status=200):
Expand Down Expand Up @@ -810,15 +798,6 @@ def do_POST(self):
else:
self._json({"choices": [{"message": {"content": GHOST_TEXT}}]})

# The AUTO safety check (ADR 0016 §6): a fresh-context user message
# whose prose is the vetter's static instruction.
@staticmethod
def is_vet_request(req):
users = [m for m in req.get("messages", []) if m.get("role") == "user"]
if not users:
return False
return "safety check" in content_text(users[-1].get("content", ""))

def _sse(self, req):
self.send_response(200)
self.send_header("Content-Type", "text/event-stream")
Expand Down Expand Up @@ -846,9 +825,7 @@ def _sse(self, req):
time.sleep(0.09)
reasoning_deltas += 1
# This connection's stock number (#125): the reply closes with it,
# making the exchange's end a unique needle. A future episode
# dispatch must pop EPISODES HERE and take no number for episode
# replies — their text is already unique.
# making the exchange's end a unique needle.
n = next(MockHandler.STOCK_REPLIES)
for token in REPLY_TOKENS[:-1] + stock_closing_tokens(n):
frame = {"choices": [{"delta": {"content": token + " "}}]}
Expand Down Expand Up @@ -1872,139 +1849,6 @@ def scenario_drop_refusal(self):
assert "rest" not in self.text(), \
"the refusal still streamed a reply — a request went out"

# -- the interaction surface (issues #133/#134/#135, ADR 0028) --------

# Enqueue the scripted episodes the mock pops one per chat connection
# (a tool-calling episode, then the post-answer reply).
def script(self, *episodes):
MockHandler.EPISODES.extend(episodes)

UP, DOWN, SPACE, ENTER, ESC = (b"\x1b[A", b"\x1b[B", b" ", b"\r", b"\x1b")

# The approval as a choice question (#133): the composer is replaced by
# "Approve write_file?" with [Approve, Reject] — the preview stays the
# evidence in the log — arrows move, Enter answers, the tool runs and
# the result rows in. Each act asserted before recording.
def scenario_approve_choice(self):
self.script(
("tools", [("call_w", "write_file",
json.dumps({"path": "tmp/capture-approve.txt",
"content": "approved by the choice widget\n"}))]),
("text", "File written."))
self.type("add the capture fixture file")
self.send(b"\r")
self.wait_for("Approve write_file?", timeout=10)
assert self.find_row("❯ Approve") is not None, \
"the approval's Approve option never rendered"
assert self.find_row(" Reject") is not None, \
"the approval's Reject option never rendered"
assert self.find_row("↑↓ move · ⏎ answer · esc reject") is not None, \
"the approval's hint row is clipped — the keys are unreadable"
self.wait_quiet()
# Move to Reject and back — the motion the GIF must show — then
# approve.
self.send(self.DOWN)
time.sleep(0.4)
assert self.find_row("❯ Reject") is not None, \
"Down never moved the approval's selection to Reject"
self.send(self.UP)
time.sleep(0.4)
self.send(self.ENTER)
self.wait_for("write_file → ok", timeout=10)
self.wait_for("File written.", timeout=10)
self.wait_quiet()

# The model's question, single-select (#134): the same widget asks the
# user, Enter returns the answer as the tool result, and the model acts
# on it in the next episode's reply.
def scenario_question_single(self):
self.script(
("tools", [("call_q", "question", json.dumps({
"question": "Which database should the config target?",
"options": [{"label": "Postgres", "description": "the sturdy default"},
{"label": "SQLite"}],
"multiple": False}))]),
("text", "SQLite it is — minimal setup, one file."))
self.type("set up the database config for me")
self.send(b"\r")
self.wait_for("Which database should the config target?", timeout=10)
assert self.find_row("❯ Postgres") is not None, \
"the question's first option never rendered selected"
assert self.find_row("SQLite") is not None, \
"the question's second option is clipped"
assert self.find_row("↑↓ move · ⏎ answer · esc dismiss") is not None, \
"the question's hint row is clipped"
self.wait_quiet()
self.send(self.DOWN)
time.sleep(0.4)
self.send(self.ENTER)
self.wait_for("question → answered: SQLite", timeout=10)
self.wait_for("SQLite it is", timeout=10)
self.wait_quiet()

# The model's question, multi-select (#134): space toggles, Enter
# answers every checked label in option order.
def scenario_question_multi(self):
self.script(
("tools", [("call_m", "question", json.dumps({
"question": "Which runtimes need setup docs?",
"options": [{"label": "Elixir"}, {"label": "Rust"},
{"label": "Python"}],
"multiple": True}))]),
("text", "Docs for Elixir and Python, coming up."))
self.type("write setup docs for the runtimes i use")
self.send(b"\r")
self.wait_for("Which runtimes need setup docs?", timeout=10)
assert self.find_row("[ ] Elixir") is not None, \
"the multi question's checkboxes never rendered"
assert self.find_row("[ ] Python") is not None, \
"the multi question's last option is clipped"
assert self.find_row("space toggle") is not None, \
"the multi question's hint row is clipped"
self.wait_quiet()
self.send(self.SPACE) # check Elixir
time.sleep(0.4)
assert self.find_row("[x] Elixir") is not None, \
"Space never checked Elixir"
self.send(self.DOWN)
time.sleep(0.3)
self.send(self.DOWN)
time.sleep(0.3)
self.send(self.SPACE) # check Python
time.sleep(0.4)
assert self.find_row("[x] Python") is not None, \
"Space never checked Python"
self.send(self.ENTER)
self.wait_for("answered: Elixir, Python", timeout=10)
self.wait_for("Docs for Elixir and Python", timeout=10)
self.wait_quiet()

# AUTO never asks (#135): the dangerous call auto-rejects with the
# safety check's reason — the preview receipt, the [auto: rejected]
# entry, the tool-result row — and NO approval ever renders. Still
# capture (the transcript record, not motion).
def scenario_auto_reject(self):
self.script(
("tools", [("call_x", "execute_command",
json.dumps({"command": "rm -rf /tmp/precious"}))]),
("text", "Understood — that command was refused. I will list what "
"is safe to delete instead."))
self.type("/mode auto")
self.send(b"\r")
self.wait_for("[mode: Build → Auto]", timeout=10)
self.type("clean up the precious directory")
self.send(b"\r")
self.wait_for("[auto: rejected execute_command — deletes files "
"outside the project]", timeout=15)
self.wait_for("auto-rejected: deletes files outside the project",
timeout=10)
self.wait_for("Understood — that command", timeout=10)
assert self.find_row("Approve execute_command?") is None, \
"AUTO prompted an approval — the leak the issue reports"
assert self.find_row("Approve this change?") is None, \
"AUTO armed the confirming activity row"
self.wait_quiet()

# The Tab-complete flow (issue #105; ADR 0020 amendment): "/mod" + Tab
# completes to the highlighted head "/mode " (registry order — mode
# precedes model; popup closes), then "/c" + down + Tab completes the
Expand Down Expand Up @@ -2821,12 +2665,6 @@ def assert_claims():
app.scenario_attach_cjk()
elif args.scenario == "drop-refusal":
app.scenario_drop_refusal()
elif args.scenario == "auto-reject":
# The transcript record (#135): preview receipt, the
# [auto: rejected] entry with the check's reason, the
# auto-rejected result row, the model's adaptation — and
# the scenario itself asserts NO approval ever rendered.
app.scenario_auto_reject()
elif args.scenario == "model-switch":
app.scenario_model_switch()
elif args.scenario == "resume-picker":
Expand Down Expand Up @@ -3008,28 +2846,6 @@ def assert_claims():
app.scenario_drop_image()
time.sleep(0.5)
app.gif(args.out, crop_rows=24, min_span=1.0)
elif args.scenario == "approve-choice":
# The approval in motion (#133): the composer is replaced by
# the choice question, Down/Up move the selection, Enter
# approves and the tool's result rows in. Each act asserted by
# the scenario before the GIF is written. Full frame.
app._mark = time.monotonic()
app.scenario_approve_choice()
time.sleep(0.5)
app.gif(args.out, crop_rows=24, min_span=1.0)
elif args.scenario == "question-single":
# The model's question in motion, single-select (#134).
app._mark = time.monotonic()
app.scenario_question_single()
time.sleep(0.5)
app.gif(args.out, crop_rows=24, min_span=1.0)
elif args.scenario == "question-multi":
# The multi-select flow (#134): space toggles two answers,
# Enter returns both in option order, the model acts on them.
app._mark = time.monotonic()
app.scenario_question_multi()
time.sleep(0.5)
app.gif(args.out, crop_rows=24, min_span=1.0)
elif args.scenario == "growing-session":
# Mark BEFORE the first send: the GIF covers both turns — the
# span appears with the first turn's persist and grows with the
Expand Down
13 changes: 0 additions & 13 deletions test/dash_float/rig_needle_test.exs
Original file line number Diff line number Diff line change
Expand Up @@ -52,19 +52,6 @@ defmodule DashFloat.RigNeedleTest do
"of 1048576",
"window assumed",
"DASHFLOAT_CONTEXT_TOKENS",
"write_file → ok",
"File written.",
"Approve write_file?",
"Which database should the config target?",
"SQLite it is",
"question → answered: SQLite",
"Which runtimes need setup docs?",
"answered: Elixir, Python",
"Docs for Elixir and Python",
"[mode: Build → Auto]",
"auto-rejected: deletes files outside the project",
"[auto: rejected execute_command — deletes files ",
"Understood — that command",
"[session titled: {TITLE_TEXT}]",
"[session titled: {TITLE_FREE_TITLE}]",
"[session titled: {TITLE_INLINE_TITLE}]"
Expand Down
Loading