B1: a slot save due at a boundary completes before any State.Capture or FLYSIM01 export there;
the ported milestone archive holds the post-capture ratchet and slot where legacy holds the
pre-capture ones, declared in legacy-gameboy-v1 sections 4 and 16 and step-v1 sections 3 and 6.
N1 the exact restore transient (blocked window from 0 ms, restored winner reported blocked on the
first decode); N2 the decoder-config form; N3 worker status and lost-reply resolution for the
extension methods; N4 boundary actions and captures in the step-v1 section 8 trace; N5 a warm-up
override is another profile; N6 sugar refused until the first commit after a restore.
Review round 1, N2. The gameboy-decoder-config-v1 form (channels as ordered arrays), vectors for
raw mode and the Pokemon Red macro group computed by flysim's legacy_profile_identity test from
gameboy_decoder_config_with_macros and reproduced by @flybrain/session-types from the oracle's
gameboyDecoderConfig. The example composition now carries the real macros-mode digest and
channel set.
Review round 1, B1 and N4. TraceBehaviour.boundaryActions records the slot saves and the rollback
at the reached boundary in application order; TraceOperational.captures records each capture with
the number of boundary actions before it; TransitionTrace refuses a capture taken before the
boundary's slot saves. Both languages, fixtures for the rule, synthetic coordinator records both
lists empty. contractDigest moves to a56e25e6.
The operator decided on 2026-09-23 to port the live fly onto the session framework in full.
New contract legacy-gameboy-v1: the profile gameboy-legacy-fafb-v783-v1, the proof that the
legacy f64 frame clock equals the rational one, the step-by-step placement of step_frame in
lockstep-v1, the readout context (location allowed and declared), the channels decision, the
memory-image inspection and ROM AssetRef, the environment (one no-button setup frame, u8->f32
audio, DC blocker at the edge, gameboy-slots-v1), the pokered-macros-v1 executor as one object
with its task, the legacy-ratchet-rollback-v1 policy, the composition digest carrying decoder
and macro-channel configuration, legacy-transient-reset restore semantics, sugar admission with
a one-commit lag, and FLYSIM01 as format of record until RETIRE-01. PROF-02b is a stub.
Dated amendments, each citing the decision, where earlier text kept the legacy loop outside
lockstep or had no place for it: workers-v1 (telemetry, Initialize, executor, episode request
kind, admission, new section 7 extension methods), step-v1 (rollback edge, Phase B/C, clock,
episode policy, sugar lag), state-media-v1 (audio, memory-image retention, restore semantics,
format of record, section 7 ratchet), README section 4, implementation.md (AGENT-01 and ENV-01
unblocked), the MaleCNS backlog (FOUNDATION-02 split, RUNTIME-01 contract) and analysis (5.2,
5.4), and readout.md (where the location comes from).
PROF-02a and RT-01a, machine-readable half, per the operator's port decisions of 2026-09-23.
Generic (in the session schema set, so contractDigest moves):
- AgentTelemetry.stimulusRemainingMs (number|null): sugar admission reads the pulse from the
last commit.
- EpisodeRequest.kind is terminal | rollback.
- Environment.SaveSlot / Environment.RestoreSlot (capability gameboy-slots-v1) and
Agent.Rollback (capability legacy-ratchet-rollback-v1) payloads, with their scope checks;
maxSlots 4.
Legacy Game Boy (module gameboy, digested apart from contractDigest):
- registered payload schemas gameboy-readout-context-v1, gameboy-channels-v1,
gameboy-joypad-v1, gameboy-memory-inspection-v1, legacy-ratchet-rollback-v1, each SchemaRef
digest over its canonical declaration;
- the one legacy profile gameboy-legacy-fafb-v783-v1 embedding today's schema-1 fingerprint,
lif-1ms-f64-v2 and fly-kc-mbon-rstdp-v2, with its AssetRef digest;
- the composition declaration carrying the decoder and macro-channel configuration, the
executor pokered-macros-v1, gameboy-slots-v1, legacy-ratchet-rollback-v1,
legacy-transient-reset and FLYSIM01 as format of record, cross-checked against the FLYSIM01
compatibility string.
Fixtures regenerated by update_fixtures (new derived gameboy-legacy.json); valid/invalid cases
for every new type in both languages; the frame clock proven identical to the legacy f64
accumulator. flysim gains only a test (and a dev-dependency) that recomputes the pinned
fingerprint, versions, frame size, warm-up, clock and button order. The synthetic fly-session
agent reports stimulusRemainingMs null and its task names kind terminal; no runtime change.
Row 41 counted `YES` starts per map, which cannot tell one conversation from
another; the Pewter Gym stall was 64 `YES` and 62 `NO` in ten brain minutes at
one person. So the harness counts an answer to a box it can read as a choice,
charged to whatever the fly is facing -- `palette::facing_target`, which is
`TALK`'s own precondition and the same reading its ledger entry uses.
`the_fly_leaves_the_pewter_gym_guides_ring_from_the_rung_ten_checkpoint`: the
fly leaves map 54 on frame 1,337 against never in twenty brain minutes, spends
27 macros in the gym (one lap of the ring), answers the guide's prompt **once**
against a bound of four, re-opens the ring **never**, and deals `NEXT` on no pad
while a readable prompt is open.
The rung is printed and not asserted: neither this run nor either arm of the
trap hunt reaches rung 11 from this checkpoint, and why the fly does not walk
back through the gym door is row 12.5's ground and row 54's rather than this
row's. Named in the residuals.
The survey whole, both arms of the trap hunt, the ROM run, and the residuals --
including the two this row does not work: neither arm reaches rung 11 from this
checkpoint, and the watchdog's check 10 cannot see a loop of exactly four
distinct macros.
The other half of row 56. The Pewter Gym guide's conversation is fifty-two boxes
long and a third of the presses that walk it are `NO`, whose B advances a plain
box exactly as `NEXT`'s A does. Every one of them cleared the pending `TALK`, so
the talked ledger never learned the conversation had happened, `TALK` stayed on
the overworld pad, and an A press at him reopened the whole ring: thirty brain
minutes of scene `dialog` with no walk macro dealt.
12.4's rule -- "the fly said no, so the thing is still on offer" -- is about a
declined *offer*, and which of the two a `NO` was is decided where it can be
seen: by whether the box closes on it. So the decision moves to the frame the
text goes away, which is where the talked entry is written anyway, and the
reading is `pending_answer`: armed only by an answer to a prompt this crate can
read, alive for one hold. A declining `NO` still standing there is a `NO` the
box closed on; anything else is a conversation walked through to its end.
Row 41 read the border at (11, 6)-(19, 11) because that is where the Pokemon
Center's script puts it, and named the limit in its own residual: "Red places a
two-option menu where the script asking for it says, so a prompt drawn elsewhere
reads false and keeps the pad it had". The Pewter Gym guide draws the same menu
at (14, 7)-(19, 11) with the cursor at column 15. Surveyed over 260 presses from
the live checkpoint: the box is drawn on 10 frames, `yes_no_prompt` answered
false on all 260, and `wTextBoxID` read `TWO_OPTION_MENU` on exactly the 10.
So the pad was `NEXT, YES, NO` on a box that was a choice -- two channels for
one A press, which is 12.10's forbidden pair -- and the reopened-prompt
exclusion never armed, because it only judges an answer to a prompt this crate
can read.
The screen half is now the border drawn **around the cursor the game parked in
it**. One fact about `DisplayTwoOptionMenu` rather than about any script: the
cursor goes in the box's first interior column, so the left edge is one column
to its left, in both boxes surveyed. The top is not fixed -- the nurse's box
begins two rows above the first item and the guide's one -- so the top is found
and the figure is then read whole, as `waiting` and the move list are.
Row 56 is a gym's dialog for thirty brain minutes with YES, NO and NEXT dealt
in equal thirds, so the question is which box each press is answering and
whether the seam can see it at all. `state::drawn_boxes` asks the screen
instead of a pinned rectangle: every complete `TextBoxBorder` on the frame.
`FLY_PROBE_CATCH=dialog` walks the conversation one raw pulse at a time and
prints both readings of every frame side by side, with the pad the palette
deals for it, and tallies how the candidate readings separate the frames.
`FLY_PROBE_ANSWER` picks the button a frame with a box on it is answered with,
so the two arms of a choice can be walked separately.
Three conflicts, all of them two reviews appending in the same place, all
resolved by keeping both sides: `macros.md` (row 50 keeps section 12.18, row 55
becomes 12.19), `macros-traps.md` (both trap rows, both residual pairs, both arm
sections) and `scene_probe.rs` (both survey modes, `accept` and `shop`). Every
source file auto-merged.
The two rows compose on the cartridge: from the mart checkpoint the fly leaves
the counter, walks the town, and no `MOVE n` blocks any more, so the longest
chain of one macro blocked on an unchanged frame is 1 in any scene.
`macros.md` section 12.18 is the review; `macros-wram.md` section 7.1 corrects
the address table, where `wListMenuID` was said to be cleared by every text
display and an item's position in the stock was said to be its cursor index;
`macros-traps.md` carries row 55, its two arms and its ROM run, and two new
residuals -- `wListScrollOffset` is not a pinned address, and a mart that has
never drawn a buy list reads `Unknown` rather than `Shop`.
The press survey, twenty brain minutes before and after on both checkpoints,
and the ROM-gated forest run.
The ethos check's 'fewer flagged windows, more distinct tiles' holds on the
tiles on both arms (296 to 430, 163 to 184) and does NOT hold on the windows
(33/73 to 70/73, 68/73 to 73/73). After the fix the fly spends three quarters
of each run inside battles it is actually fighting, and the hunt's tile rule
flags a fighting fly exactly as hard as a stuck one. Reported rather than
smoothed.
take_events kept returning Option<EventBatchView> and defaulting through the
question-mark operator, so a batch missing a field read as end of stream and the
ConsumerEvents enum added in the previous round described nothing. It returns
Batch or Unreadable now, and a test publishes a batch with no droppedBefore,
asserts it is reported as unreadable naming the field, and asserts the next real
batch still reads.
The checkpoint-envelope-v1 amendment cites the rule that lets a required
manifest field land with envelopeVersion still 1 while no production file
exists.
Dropping the length check left nothing checking the reply against the
request at all: AcknowledgeResult::validate_against existed with no
caller, so a worker could acknowledge ids this session never asked
about. That is the other half of the rule the README states. A short
list is the worker reporting what it released and is accepted; an id
from outside the request is the worker reporting about someone else's
cache and is refused, named, before any mutation.
The bootstrap-path regression is now covered too. The earlier test calls
acknowledge_replies directly, which guards the check where it lives but
not where it lived, so a length check put back into acknowledge_lifecycle
left it green. The duplicate_lifecycle_acknowledge injection releases the
ids first, out of sight, so the call that method makes and checks is
already the second one -- the shape the section 6 resolution produces.
Verified by putting the old check back: three tests fail with it, none
without.
Also: last_resolution_attempts is cleared with last_resolution, so a
resolution ending before its first attempt no longer reports the
previous count; the guard half of the bound test asserts the fence like
the budget half; the attempts assertion checks a real bound rather than
u32::MAX; and two dead Instant bindings are gone.
The index an agent attested to at Agent.Initialize joins its compatibility
identity and its checkpoint manifest row, so a replacement fly that built
another graph -- the same dataset, the same neuron count, another index --
cannot install a checkpoint taken under the first one. It was accepted before,
because agent_compatibility digested a dataset digest recomputed from a free
function rather than what the worker attested to, and the restored composition
was then published under its predecessor's indexDigest.
The refusal names what the worker is, not just that two digests differ. Dated
amendment to checkpoint-envelope-v1's agents row; no wire type and no schema
text change, so the contract digest is unchanged.
AgentSlot.telemetry was written only by the Agent.Initialize handler, so every
CommittedSnapshot carried warm-up telemetry labelled as boundary k while each
AgentCommitResult.telemetry was validated and dropped. The commit result is now
stored on the slot beside the committed step, and the media test asserts that
published telemetry advances across boundaries instead of merely being nonzero.
A refused snapshot is recorded and sequenced like a refused descriptor revision,
because the repair path exists for the consumer that did not receive it; the
sequence advances with the value rather than with the delivery, so two snapshots
can never share one. The query service counts an answer it could not deliver
rather than discarding the result, and an unreadable event batch is distinct
from the end of the stream.
From the checkpoint the stream looped in: the fly is out of the shop scene and
off the map inside twenty brain minutes, no macro reports `blocked` more than
three times in a row in the shop with the map, the tile, the wallet and the scene
unchanged, and the shop never deals an empty pad.
The chain is recorded per scene as well as overall, because the battle scene has
one of its own and it belongs to another review.
A mart's buy list scrolls. The cursor walks rows 0, 1, 2 and then the window
moves under it, so an item's position in the counter's stock is its cursor index
only for the first three entries; the offset that would name the rest is not a
pinned address. Pewter's counter carries seven items and the Antidote is its
fourth, so `BUY ANTIDOTE` there aimed a cursor step above the list's own max and
reported `blocked` on its first frame, having pressed nothing -- and a purchase
has no walk target, so nothing was recorded and the button came back on the next
hold.
One accessor answers "which index, if any" for both the pad and the plan, and the
clerk talking is refused through the seam rather than here.
`wListMenuID` keeps PRICEDITEMLISTMENU for a whole mart visit: the clerk's text
is printed from inside the mart's own routine and never goes through the display
that clears it. So every frame of the Pewter mart read "the priced buy list",
including the "Here you are! Thank you!" box the stream was looking at, whose
leftover cursor bytes belong to a two-option box.
Which screen is up is now read from the figure the game draws, the same
construction the dialogue box's `waiting` test and the YES/NO prompt already
make: a full-width box waiting for a press is the clerk (`ShopScreen::Talking`),
the item window drawn is the buy list, and the item window blank is the counter
menu.
`FLY_PROBE_CATCH=shop` drives a `BUY ...` from a checkpoint the stream looped in
and prints, per frame, what the seam reads at the counter, the halves of every
purchase's precondition, the outcome of the macro three times over, and what an
A press at the counter really opens. `screen_text` decodes the tile buffer,
because the question is which figure the game drew and the bytes that would name
it are not rewritten between two of the mart's screens.
The flybus session_over_one_router failure my workspace runs were
counting is already fixed on main: the coalescing branch forces the
coalescing deterministically instead of asserting that a slow consumer
must skip. Merging before the runs so they measure a tree that exists.
One conflict, in the crate README, and it was two sections both newly
added at the same anchor rather than two versions of one thing. Both are
kept: STATE-01's checkpoint and recovery section, then the guidance on
which replies permit a subset, which stays immediately above the
contract-narrowing section where someone adding a check will meet it.
coordinator.rs and the process tests auto-merged. Checked rather than
assumed: the acknowledge equality check is still gone, acknowledge_replies
and last_resolution_attempts are present, STATE-01's capture, durable
and rebase surfaces are present, and both test changes survived.
Both death rows killed after a fixed sleep, so under load the kill could
land before the call was dispatched. The bus then reports not-dispatched
and MutationCertainty::None, which is correct -- the participant never
received anything -- while the test demanded unknown. A full-workspace
run caught it: "left: None, right: None" at the certainty assertion.
kill_once_it_is_working polls the victim's own Worker.Status until it is
provably inside the operation before killing: the agent until it is
Preparing with an active request id, the world until it has recorded the
batch, which the arena does before its injected delay. Dispatch has then
demonstrably happened and unknown is the only correct certainty.
The wall-clock boundedness assertions are gone with them. The suite's
own `within` is the bound, and the victim is five seconds slow against
its twenty, so returning at all is the claim.
A group restore re-establishes a committed boundary this epoch did not run a
transition into, so its snapshot carries no decision and no controls for any
agent, and a fresh epoch is a new compositionDigest, so the descriptor takes the
next revision rather than republishing revision 1 with different contents.
CommittedSnapshot's rule becomes: null at boundary 0 and at an installed
boundary, always together, and for every agent or none -- a snapshot where one
fly acted and another did not would be two boundaries in one value. Dated
amendment to publishing-v1 section 3, with the schema set, the fixtures and the
TypeScript package moved together and the digest regenerated.
The first pass of the press survey counted 187 refusals that were its own held
button: JoypadLowSensitivity acts on a key's edge, so a direction the fly was
already holding produced no press. With the pulse releasing first, the reading
is 264 honoured of 3,102 by the cursor bytes alone and 231 of 231 by the bytes
and the box.
The sweep behind this branch found one check demanding an exact match
where the contract permits a short answer, and four that were right to
demand one. The question that separates them belongs where the next
check gets written, not only in a run report: is the far side reporting
what it did, or being held to a requirement?
Worker.Acknowledge is the only reply of the first kind here, because
ipc-v1 section 5 makes it idempotent. The commit and batch checks are
the second kind and are named so nobody loosens them later in the name
of tolerance; they are what make a partial commit and an incomplete
batch fail.
The short-list acknowledgment is a contract rule, so it gets a test that
says so rather than one that depends on a worker being slow.
acknowledge_replies carries the ipc-v1 section 5 sentence in its doc
comment and returns what the worker actually released;
acknowledge_lifecycle calls it, so bootstrap and the test exercise the
same path.
an_acknowledge_that_releases_nothing_is_not_a_failure drives the case
directly, once per execution mode: bootstrap releases every lifecycle
reply, the test asks for those ids again, the worker ignores them and
releases nothing, and the coordinator must accept the empty list, stay
unfenced, stay at its boundary and still play the next transition. It
fails if the equality check returns.
Row 50's two numbers, ROM-gated from the forest checkpoint. Before, on main:
MOVE n 940 starts and 838 blocked, 11 battles entered and 10 ended, worst 503
macros, median 48. After: 51 starts and 0 blocked, 13 entered and 13 ended,
worst 283, median 43, and the fly leaves the forest north through the gate.
The blocked share is the assertion the row is about; the median is what it
buys, and the worst battle is a tail rather than the run.
The bound test failed two runs in thirty under load, and not on the
bound it was testing. Both failures were bootstrap: "a worker did not
acknowledge every lifecycle reply".
ipc-v1 section 5 says already released or unknown ids are ignored, and
the contract type already holds the acknowledged list to a subset of
the request. So the second Acknowledge of the same ids answers with an
empty list by design -- and the section 6 resolution produces exactly
that second Acknowledge whenever the first reply is slower than the
probe. Demanding the whole list back turned a safe, contract-sanctioned
retry into a failed epoch, which is a defect in the coordinator rather
than in the test: a slow lifecycle reply would do it to a real session
too.
The test made itself easy to hit by installing a fifty-millisecond
probe before bootstrap, so bootstrap's own lifecycle calls ran under a
budget meant for the step under test. It now bootstraps at ordinary
deadlines and tightens them afterwards.
The two bounds are also separated by construction rather than by clock.
Each half puts the bound it is not testing out of reach -- u32::MAX
attempts against a fifth of a second, three attempts against an hour --
so no scheduling delay can flip which one fires, and the silent
participant is ten minutes slow against a twenty-second test timeout,
so returning at all proves a bound ended it. The wall-clock assertion
is gone and the attempt count is asserted instead, which
last_resolution_attempts now records. One agent per composition, so the
participant the failure names is not a race either.
An undeclared stimulus kind is refused before the model is touched, proved by an
injection through Agent.Commit rather than by a unit call, so the declaration a
descriptor publishes is the thing the worker enforces.
The publishing-v1 section 2 amendment now says in its own words that
Session.GetDescriptor and Session.GetSnapshot are internal and provisional names,
which the later public v2 step may rename or supersede.
Review follow-up on the checkpoint store.
A dropped reply channel and an expired caller budget were both reported as
ReplyLost. They are different facts -- the first means the write is over and its
outcome did not reach here, the second means the save is still going -- so they are
now separate outcomes, and the durable wait has its own budget rather than borrowing
the one that bounds a call to a participant. Both still leave durable metadata where
it was, and for both the resolution asks the store about the same checkpoint.
The writer's two bounds refuse at different moments and the comment claimed
otherwise: the outstanding-capture bound is taken before a capture is requested, and
the byte budget cannot be, because a capture's size is not known until it exists. The
byte check, the decision and the change to the byte total are now one critical
section, the peak is sampled after a superseded job's bytes are gone, and the writer's
own bookkeeping is over a type that holds only the outcomes a writer can produce.
The manifest's coordinator.eventWatermarks is {lastSourceStep, issued}; the fixture
illustrated {lastEventId, lastOrdinal}, and the illustration is what changed, because
an event id is derived from the epoch and cannot be compared across the restore that
gives the session a new one.
checkpoint-envelope-v1 section 3 also now says, under the same dated amendment, that
a required-manifest-field change must bump envelopeVersion once production files
exist: contractDigest is taken over the schema set and does not cover this manifest,
so the envelope version is the only thing that can carry such a change.
