# session attachment

Session creation and user turn admission use durable [operation receipts](operations.md).
An accepted submission survives a daemon restart while it waits for transcript
consumption. A retry with the same operation ID returns the original admission.

The CLI app package owns session ID resolution. Full lowercase UUIDv7 IDs
address roots and children directly. Shortened IDs match only root sessions,
with an exact match taking precedence over prefixes. A missing full ID returns
the same no-match error as a missing prefix; resume never creates a replacement.
Opening reads the root list and settings once, then captures the selected session
once and adds a child to its initial list if needed.

opening a session attaches the cli to its event stream and independently
reads `GET /sessions/{session_id}?tail=0`. replayed history does not establish the
session's live phase; status and live events do.

an idle status is enough to enable the composer, including a persisted
`interrupted` session after a daemon restart. a missing or unrecognized
phase is display metadata, not a failed connection. before live status is
known, enter preserves the draft; superseded status replies cannot unblock
it. starting a new turn does not require `/compact` to change the old phase.

## Live status without a poll

`GET /sessions/{session_id}?view=status` answers the live facts alone: `id`,
`cursor`, `status`, `kernel`, `usage`, and `glances`. Its source is the session
actor's own memory plus its kernel's port owner, never configuration, history,
or a composition read; the one exception is a session whose sidebars no stream
has published yet, where the handler observes the composition for that answer.
Tail and replay parameters belong to the event stream and are refused here.

Most of what that read covers also arrives on the stream, so a watching client
needs no cadence of its own:

- `kernel` carries `{"kernel": Kernel}`, the same object the session resource
  carries. `albedo_python` runs the closure the session registered with
  `kernel.watch` whenever a frame changes the job facts it reports (a job
  starting or ending, a remote job listing, a live cell, the kernel attaching
  or going stale), and the session publishes `session_namespace.kernel_facts`
  only when they differ from what it last sent. Nothing observes the kernel on
  a timer.
- `glances` carries `{"glances": [Glance]}`, the extension sidebars. Another
  session writing the work ledger or a vent publishes nothing here, so while at
  least one client watches, the session re-reads the sidebars every 10 seconds
  (`session_state.glance_ms`) and publishes only a change. The observation runs
  on a spawned process because it queues behind kernel boots, and the last
  watcher leaving cancels the timer.
- The turn's own phase keeps arriving as `status`, usage as `usage`, and a
  reset batch re-seeds all of it for a fresh attachment.

A daemon from before these events never sends them, so the CLI keeps one safety
read (the status view) every 60 seconds while its stream is healthy, plus the
read it makes when it attaches or the terminal regains focus.

## Running observations

`daemon/albedo_bus.erl` owns the ephemeral running cache. Only active turns
retain a `{SessionId, OwnerPid}` record. An idle publication removes the
calling actor's exact record, and both registry owner-death handling and the
progress owner monitor remove the stopped actor's record. Delayed cleanup
cannot remove a replacement actor's active mark. Missing records and dead
owners read as idle, so the registry can check activity without calling a busy
session. The table lasts for the daemon lifetime; it retains no idle or deleted
session IDs after owner cleanup. Changes to this cache require a fresh daemon
lifetime rather than hot-loading code.

## Multiplexed browser streams

`GET /events` upgrades to a WebSocket through the ordinary bearer-token and origin admission path. One socket process owns up to 64 follows; it never spawns an SSE writer per follow. The browser proxy supplies the bearer token.

Client frames select a key: `{key,type:"session",session_id,tail:0}` with optional `after_generation` and `after_seq`, `{key,type:"sessions",scope:"roots"}`, or `{key,type:"stop"}`. Subscribing an existing key replaces it. Keys contain 1 to 128 UTF-8 bytes; commands are at most 4096 bytes. Session replay parameters use the SSE validator. The response `{key,batch:...}` embeds the same SSE batch, including resets, snapshots and failures. `{key,end:true}` ends one follow, including an explicit stop, deleted session, failure or collection overflow. Invalid commands and the follow cap return `{key,failure:{code,detail}}`. A malformed command without a usable key has an empty key. `{type:"ping"}` receives `{type:"pong"}`; the socket also sends a pong every 15 seconds.

`http_session_stream` owns replay reads, batch encoding and keyed watch/ack/close helpers for both transports. `http_session_collection.Follow` owns the collection filter, bounded bus drain, batch encoding and rearming for both. The socket sends only from its own process and acknowledges a cursor only after a successful frame. Each frame is at most the HTTP response limit, with envelope room reserved in the shared encoders. Producer wakes remain coalesced; stale wakes from replaced keys cannot rearm the new follow. Keyed session watches allow several follows of the same session on one socket. Unsubscribe removes watches and monitors; socket closure closes every follow, and producer monitors also clean up a dead owner.

## Stream failures and recovery

Internal checkpoint and provider-started signals are suppressed before stream
encoding and do not allocate SSE sequence numbers.

Session streams require the `text/event-stream` content type, zstd-coded when
the request accepts it ([encoding](../docs/http-api-design.md)): `http_stream`
flushes the coder after every frame and ends the frame before a deliberate
stop. A zstd context works only in the process that created it, so the stream
process creates its own writer. Frames are coded with a 64 KiB window, which
is the decoder memory a client holds for the life of the stream
([encoding](../docs/http-api-design.md)). Each batch carries
a nonempty `generation`, a nonnegative integer `cursor`, and an `events` array.
The generation identifies one lifetime of the session actor. Reconnect requests
send both `after_generation` and `after_seq`; an initial request omits both.
A missing or different generation, or a sequence outside the replay window,
returns a durable transcript reset. Empty keepalive batches carry the current
pair too.

Attachment requires `session_replay` capability version 2. Reset snapshots
include `active_output`: unfinished text and thinking captured with the same
actor cursor as the durable history watermark. The session actor owns this
projection. Do not reconstruct it from the replay ring or bounded activity
previews, which may have already discarded the beginning of a response.

The projection retains at most 64 KiB of raw inline text and emits at most
64 KiB of encoded inline content across its segments. Larger output spills to
temporary files with 16 KiB buffered writes. Capture flushes pending writes
before returning references. Each reference fixes the byte cutoff at capture;
later deltas cannot extend that prefix. Full-content pages use
`/sessions/{session_id}/active-output/{content_id}` and retain the original
`snapshot` query while following `next` tokens.

The CLI stages complete prefix hydration before replacing its transcript,
then restores history and active output before processing later events.
Provisional chunks retain message identity even after the live display buffer
settles them. Canonical assistant and thinking publication precedes
`committed.replaces_live_ids`, which identifies matching provisional chunks.
Model commits set `replaces_all_live` to retire all provisional output, including
fragments absent from the bounded projection after a storage failure.
Retry identities never reuse a failed attempt's content. What a turn streamed
but never finished is committed rather than dropped; see [unfinished
turns](#unfinished-turns).

Issued spill references live for 15 minutes across commit and retry. Existing
maintenance removes expired and orphaned files; session deletion revokes them.
Spill storage is bounded to 64 MiB per run, 1 GiB across retained files, and
4096 retained files. A projection has at most 256 segments. Spill directories
use mode `0700`; content and metadata use `0600`. Owner identity includes the
VM incarnation so reused PIDs cannot retain orphaned files after restart.
Storage failure makes backfill explicitly unavailable rather than returning a
truncated successful snapshot. It must not abort generation or discard a
durable commit. Expired references return `410 active_output_expired` and cause
the client to capture a new snapshot. Do not add compatibility readers for
snapshots without active output.

Text and thinking deltas reach subscribers at most once a frame (16 ms):
`session_state.emit` publishes a delta after a quiet frame at once and holds
later ones until the frame ends, merging neighbours of one message into one
event. Any other event, a commit, a retry, and the end of a turn publish the
held deltas first, so the stream keeps its order and active output never lags
a commit. The replay window keeps at most 256 events and less than one
response's bytes, so any replay fits a single batch.

Session subscribers receive coalesced wake notifications. Event payloads stay
in the session's existing replay buffer; a successful frame acknowledges its
cursor before another wake is issued. A subscriber's death removes its watch
without requiring another event. A session that stops (closed at shutdown or
deleted) wakes its subscribers, and a stream whose session stopped or did not
answer ends with a clean EOF rather than a failure batch: the daemon closes
sessions before it stops listening, so that is a restart or a deletion, and the
client's reconnect learns which.

The CLI validates the whole batch before consuming it and saves both cursor
values only after every callback succeeds. Initial attachment and a generation
change require a leading reset. Within one generation, the sequence cannot
decrease without a reset. A generation change with a reset is normal recovery.
A change without a reset is a protocol failure. The reset replaces visible
history and clears live tool progress. Reset batches include `current_progress` in the captured session snapshot. The CLI delivers that snapshot after history as
live progress, so attaching during a tool call restores its current tool name,
phase, and bounded code preview. Older-history navigation uses transcript row IDs.
Unknown event kinds are ignored.
Missing kinds and malformed known events are protocol failures.

Transport failures, ordinary EOF, HTTP 408/429, and server errors reconnect with
a delay starting at 500 milliseconds and increasing to at most five seconds.
EOF and transient failures retain the saved pair. Cancellation ends the
subscription quietly and clears the pair, as does explicit protocol recovery.
Other HTTP refusals, explicit SSE failures, and consumer callback failures end
the subscription with a visible notice, except an SSE failure coded
`daemon_unavailable` or `daemon_stopping`, which older daemons send while they
shut down and which reconnects like a server error. A terminal failure also stops status
polling for that attachment.

The TUI attempts one durable transcript reset per attachment after a protocol
failure. It clears live tool progress and requires the recovery stream to
begin with a reset. A second protocol failure ends automatic recovery; reopening
the session starts a new attachment. Pending submissions and drafts remain
available, and a stream failure does not claim that model execution finished.

## Notes in the transcript

The TUI renders daemon notes as faint plain text, wrapping each line before
styling it. Never style a whole multi-line note as one Lip Gloss block: that
pads every short line to the longest line, which can span thousands of columns
in a capability update. Long notes use the same click-to-expand fold as user
messages. Folding changes only the display, not the stored or model-facing text.

## Loaded sessions and unloading

A session is loaded while its actor runs; only a loaded session has a stream
cursor. Opening it, watching its stream, or submitting to it loads it. Reading
`GET /sessions/{session_id}/history` never does, so the session list previews
stored history instead of the session resource.

A history page carries a reference instead of any content field over the page's
inline budget. The client resolves every reference of a page in one walk of
`GET /sessions/{session_id}/history/content?entries=...`, which serves the same
parts as the single-entry route for up to 200 entries and stops on the same
byte budget, so a page of oversized tool output costs one request per byte
budget rather than one per entry. Decode each part once: the reference
resolution, the projection into events, and the tool-call argument merge all
read the decoded parts, and a decoded `json` part keeps the stored bytes as its
value rather than another JSON encoding of them.

The maintenance sweep releases an idle kernel after `ALBEDO_IDLE_SECONDS`
(default 10 minutes) and unloads a session whose last turn (`activity_at`) is
older than `ALBEDO_UNLOAD_SECONDS` (default one hour): the provider's prompt
cache has expired by then. The session refuses while a turn runs or waits, a
client watches its stream, a background job is live, or one of its children
works — runs a turn, or sits between turns on a job of its own, or has a child
that works: that child's report will wake it, and its prompt cache may be kept
warm meanwhile (robot-docs/cache-warming.md). Unloading saves the kernel's
variables like a release. The next request loads it again.

Listing sessions reads loaded sessions' summaries but does not count as
attention, so an open session list never keeps a kernel alive.

The same sweep collects every waiting process whose heap passed 1 MB
(`albedo_daemon:collect_idle`): a long-lived actor otherwise keeps the heap a
busy turn grew it to. The launcher's VM defaults (`priv/bin/albedo-daemon`)
then hand the freed carriers back to the OS instead of caching them.

The session list marks a session active (green) while it is loaded and its
last turn is under an hour old.

## asynchronous persistence failures

The dispatcher logs failures to advance delivered schedule occurrences or save mail delivery errors. Undelivered mail and unadvanced schedules remain durable and retry under the existing dispatch policy. A schedule whose submission succeeded but whose advance failed can be delivered again.

## Prompt ownership and cancellation

Inputs use `PUT /sessions/{session_id}/inputs/{input_id}` with a client-generated
UUIDv7. The same identity follows admission, transcript consumption, turn
membership, and cancellation. Accepted inputs survive restart. The
[input contract](../docs/http-api-design.md#input-admission-and-outcomes)
defines message, continuation, skill, and command variants.

`POST /sessions/{session_id}/inputs/{input_id}/cancel` returns the current input
and `cancelled`, `interrupt_requested`, `shared_running`, or `not_pending`.
A waiting cancellation commits its receipt and pending removal together. A
running turn shared with other inputs continues. Cancellation never falls back
to a session-wide interruption or affects a later unrelated turn.

The CLI uses a separate five-second cleanup deadline and reports unconfirmed
cancellation when cleanup fails. An interruption acknowledgement does not claim
that work stopped.

Live turn membership precedes worker output and expands when queued input
steers a turn. Completion follows its final output or failure. The durable input
read retains membership and terminal outcome across reconnects and restarts.
Prompt clients follow their input identity instead of inferring completion from
echo metadata or idle status. Drafts and unresolved input handles survive resets.

## Unfinished turns

A turn that stops before its output commits keeps what happened
(`session_salvage`). When the user stops it, when it fails, or when a provider
stream breaks after showing output, the session commits, in order: a result for
every call of the run that nothing answered, the text and thinking still in the
active output, and a `<system-note>` user row saying where the response ended.
The call the tool gate last let start gets "may have had effects; inspect them
before any retry", because the worker may have been killed mid-call (a cancel
kills it after 2.5 s); every other open call gets "not run". Calls whose
arguments were still streaming never ran and are named in the note.

The streamed text and thinking become one chat-shaped assistant item marked
`albedo_unfinished`, which history shows as thinking and text like any
response. No provider signed or encrypted it, so the projection never replays
it natively, and a model reads only its text, as plain assistant text. The
thinking never goes back to a model: Anthropic's `reasoning_extraction`
classifier refuses requests that carry unsigned reasoning, with an empty
response, and keeps refusing every later turn of the session. Measured on
claude-sonnet-5-5 and claude-opus-5-5, thinking sent as a `<system-note>` was
refused on 11 of 11 sessions; as `[Reasoning summary]` assistant text on some;
left out, on none. Thinking another provider produced, which reaches a new
provider as plain assistant text when a session switches, carries no label: a
`[Reasoning summary]` label only invites the classifier.

A user stop always gets the note ("The user stopped this response here."). A
failure gets one only when something was kept. A retryable provider failure
after the stream showed output does not resend the same request: the loop asks
the session to commit the output with a note to continue from it, then starts
the step again from that history, counting the attempt against the same retry
budget. A failed attempt that showed nothing, or only partial tool call
arguments, is retried unchanged.

## rewind

`POST /sessions/{session_id}/rewind` (capability `history_rewind`) cuts an idle session back to before a user message and keeps everything it removes. The module doing it is `daemon/history.gleam`, beside the fork it reuses, and it runs inside the session actor (`RewindHistory`) so no submission can land mid-cut.

One transaction does the work. It copies the whole session into a new archived session with `fork_in` (its name is `rewound: ` plus the title; `archived` is the ordinary preference, so it is a normal session that opens, forks and deletes like any other), then `remove_tail` deletes every row from the checkpoint on and what is keyed by those positions. The rules for that, in one place:

- the rows themselves, their `transcript_calls`, `transcript_quarantine`, `submission_events`, and the captured traces of the cells they ran;
- `continuation_markers` anchored at the last row kept or later, as a fork drops them;
- `provider_requests` stay for usage and lose their `seq`, which is nullable for calls with no row;
- LCM nodes whose source reaches past the cut go with their edges (`lcm_graph.truncate`), and the cursor falls back to what a surviving leaf covers;
- the preview, and the image payloads only the removed rows named (`images.release`);
- notes, rolling state and snapcompact archives are not touched: each checks its fingerprint against the history it is given, so a cut it covered is recomputed.

A table keyed by `session` and `seq` that is not in that list fails `RewindStoredStateTest` in `test/e2e/rewind_test.py`, which reads the schema; a `seq` that counts something else (a kernel link's frames) is listed there as unrelated.

When the cut succeeds the actor stops like an idle `Unload`, but ignoring its own watchers: their streams end, clients reconnect to the actor the registry starts on the next lookup, and a new actor generation means a durable reset. The registry learns of the archive through `Adopt`, which the HTTP handler sends before it answers. The Python namespace and files on disk are not rewound.

With `summary`, the handler first asks the actor to start a summary run (`StartSummary`, work `turn.Summary`). It holds the session like a background call, sends the tail to the model through the compaction summarizer (`loop.summarize_branch`, instructions in `compaction.branch_summary_instructions`), and the worker answers the waiting handler directly with the text and the last row it covered. Only then does `rewind` run, with the summary stored after the cut as a `rewind` system note, and it refuses (`history_changed`, 409) if the session has grown past that row. A summary that fails or comes back empty is a 502 `rewind_summary_failed` and nothing has changed, not even the archive.