The /update command went silent between "starting" and one final message,
gave no version summary, and required manual intervention on failure.
CommandInstaller now streams the installer script's stdout live instead of
buffering it, and parses a structured result (previous/installed versions,
per-component status) from a sentinel line. Because the reference install
script hands the daemon restart off to a detached systemd unit specifically
so restarting archied.service doesn't kill the process that requested the
update, the outcome is reported in two phases: the synchronous build/install
result comes back through Install() itself, and the post-restart health
check/rollback outcome is written by a new watchdog script and relayed by
whichever archied process boots next, via a small JSON handoff file cleared
once relayed.
Also fixes three defects an adversarial review pass found in the first
draft: the caller's request/handler context was passed straight into the
install child process, so a concurrent /restart or a browser disconnect
could SIGKILL it mid binary-copy; Telegram and the web dashboard each
tracked "install in progress" with their own private flag despite sharing
one Service, allowing two installs to race the same .prev backups; and the
watchdog reported rolled_back:true even when no backup existed to restore,
which would have told an operator to stand down while the daemon was still
broken.
internal/webui gains an unauthenticated /healthz the watchdog polls after
restarting archied.
Keep bare GitHub issue and pull-request links separate from following prose across one-shot sends, live edits, and long-message splitting. Preserve Markdown delimiters and newline boundaries while retaining plain-text fallbacks.
A daemon shutdown or /restart cancels turnCtx without any guarantee
that the goroutine owning a liveReply ever runs again to notice and
clean up after itself -- a full process exit gives it no chance at
all. Gateway now tracks every liveReply mid-turn in a registry and
Stop drains it, marking each as interrupted so nothing is left ending
in the "typing" cursor forever.
finalize and abandon now run their bodies inside a per-reply
sync.Once (the terminal field): a turn finishing naturally and a
concurrent Stop draining the same reply were reaching it from two
goroutines at once, both sending their own edit for the same message
with the outcome decided nondeterministically by which HTTP response
Telegram processed last -- including a complete real answer getting
silently overwritten by the restart marker. Only one side's body now
ever runs; the other blocks on the winner and then does nothing.
The drain itself is bounded (liveDrainTimeout, default 5s): a single
stalled Telegram API call must not hang gateway, and therefore
daemon, shutdown indefinitely. Entries still in flight past the
deadline keep running in the background rather than blocking Stop.
A liveReply created in the narrow window between Stop claiming the
registry and its own registration call is abandoned immediately
instead of being registered into (and lost from) a registry nothing
will drain again.
Regression tests cover the mutual-exclusion race directly (20
iterations of finalize and abandonRestarted racing on the same
reply, asserting exactly one outcome ever wins), the bounded drain
against a deliberately hung API call, and the post-stop registration
gap.
The finished reply is delivered by editing the message created at the
first streamed frame, and editMessageText does not change a
message's date, so a long tool-using turn's answer keeps its
turn-start timestamp and can sort before conversation that produced
it. Delete-and-resend would fix the ordering but reintroduce the
flicker/reorder streaming exists to avoid, so this documents the
accepted trade-off rather than changing behavior.
Three more P2s from the same code-review pass: the live frame's
single clamped buffer mixed tool lines and answer text chronologically,
so early tool lines silently scrolled out of the frame once the answer
grew past liveBodyMaxRunes and only reappeared at finalize (which
already led with tools separately). Tool lines are now tracked
exclusively in toolLines; a new framedText helper keeps the tool block
whole and clamps only the answer tail, used consistently by the live
frame, an abandoned frame, and the finished message.
A tool-only turn with no reply text used to finalize as a bare,
immutable tool line indistinguishable from a real (if terse) answer;
finalText now appends an explicit "(no response)" marker in that case.
A provider error used to funnel into the same clean-partial path as a
deliberate /stop, so a failed turn was indistinguishable from a
successful short one. abandonFailed appends a "turn failed" marker
before dropping the cursor; /stop still goes through the unmarked
abandon path.
Regression tests cover all three paths plus the degenerate case where
a tool block alone exceeds the frame budget, verified red-then-green
against the pre-fix code.
Three P1s from code review of PR #476: finalize sent the authoritative
reply through the same cancellable turnCtx a /stop kills, so a stop
landing mid-finalize stranded the live message with its cursor
forever; abandon had the identical hazard, saved only by its one
caller wrapping the context correctly. Both now strip cancellation
internally (context.WithoutCancel) rather than trusting callers to
remember.
Tool activity lines were composed unescaped into the same Markdown
body as the model's own reply, so a stray '*', '_', or backtick in
tool output/parameters could unbalance the whole message and take the
reply's formatting down with it. ToolCall now runs name, parameters,
and summary through bot.EscapeMarkdown before joining the line.
splitLongMessage sliced oversized lines on byte offsets, cutting
mid-rune for any multi-byte UTF-8 line crossing the 4000-byte bound
(CJK, emoji) and producing invalid UTF-8 parts. It now counts and
slices on runes throughout.
Regression tests cover all four paths, verified red-then-green
against the pre-fix code.
ai-sdk v0.1.21 wraps the cancellation cause in aborted-stream errors
("%w: %w") instead of formatting it, so errors.Is(err, context.Canceled)
is reliable again. The Telegram channel was compensating with a comment
explaining why errors.Is could not be trusted.
The code was already correct; only the comment was obsolete. It now
states the real reason for context.WithoutCancel(turnCtx): turnCtx is
dead on both stop and fault paths, so the cursor-drop edit needs a live
context. The runtime effect of the bump is that a /stop landing inside
the step loop is now logged as "chat turn stopped" rather than
"route failed".
Refs ai-sdk-0pn
Telegram streaming callbacks now enqueue cancellable rendering rather than performing network I/O while draining the authoritative model stream. Final edit failures fall back to normal split delivery.
Completed tool events now correlate inputs by call ID and expose only bounded, conservative parameter shapes. Tool-call visibility is snapshotted per reply so reloads cannot change an in-flight turn.
Chat returned only the final answer, so a wrong tool choice, a
hallucinated action, or a footgun like `systemctl stop` against
`restart` was invisible from the conversation. Only text deltas were
threaded through the turn, and tool calls executed inside Generate with
nowhere to report.
Replace the delta callback with gateway.TurnStream, which carries text
fragments and completed tool calls on one ordered path from the model
loop to the channel. The sequence is the point: a tool line that
overtakes the sentence it interrupted describes a turn that never
happened. Text-only callers use gateway.DeltaFunc and are unaffected.
A tool is reported on its result, never on its call -- the call part
carries no outcome yet, so reporting both would list every tool twice,
once with nothing to say.
Telegram renders it behind [chat.telegram] show_tool_calls, off by
default: this is debugging visibility, and it narrates every internal
step to provide it. The dashboard always shows it, being the operator's
own observability surface; its tool lines are live-only and are not
persisted with the conversation.
The Telegram live reply changes mechanism, and with it what a frame
MEANS. It used to stream through sendMessageDraft: drafts are ephemeral
30-second previews, so a tool call longer than that outlived the
preview, and successive frames were observed concatenated with their
cursors intact. It now sends one real message and edits it. An update
used to be "the next preview of the answer"; it is now "the whole
answer as it currently stands, replacing what is there". The cursor is
appended at render time and never stored, so exactly one exists and
always at the end, and a frame too long for one message is cut to its
tail rather than rejected outright.
A stopped or failed turn now edits the cursor away instead of leaving
it blinking on a message nothing will finish. That cleanup runs on a
live context because turnCtx is already dead, and it runs on both
failure branches: ai-sdk raises an aborted stream as ErrAborted with
ctx.Err() formatted in rather than wrapped, so errors.Is cannot be
trusted to tell a /stop from a fault.
Closes#467
Loader.Resolve() selects TOML/YAML files or a config directory as the
production entry point, replacing the TOML-only Overlay() as the default
path. Add internal/channels.Manager, reporting each adapter's actual
runtime lifecycle (configured/starting/running/degraded/failed) instead of
inferring health from configuration presence alone. Wire channel reload
through a per-adapter seam.
Fixes#55
Adds a human-consent gate for tools marked RequiresApproval, exposes
session management as LLM-invokable chat tools, and gives the operator
/delete to retire conversations permanently.
**Tool-approval gate (internal/tools, internal/agentexec)**
- ApprovalRequester contract lives in internal/tools (not gateway) so
agentexec can import it without pulling in the chat-adapter stack.
- ToolSetOptions.Approval wires the approver through to toolExecute.
- BuildToolSetFrom omits RequiresApproval tools when no approver is
configured; dispatch paths (invokeTool, tool_call bridge) refuse
them too, so the invariant holds at every indirection.
- The gate fails closed: only Approved/PermanentlyApproved pass;
an unknown decision value from a faulty approver does not execute.
- Tool-level timeout (2 min) returns a plain error, not a context
cancellation, so the model can report "approval timed out" instead
of the turn aborting.
- BuildApprovalDescription callback lets a tool surface the specific
resource being acted on ("Delete session \"abc123\"") rather than
showing a static tool description.
**Session tools (internal/gateware/session_tools.go)**
- session_list (idempotent), session_resume (mutating), session_title
(mutating), session_delete (mutating + RequiresApproval).
- Scoped to the current platform + channel: GetByChannel replaces
List() so a model in chat A cannot list or manipulate chat B.
- session_title validates the session belongs to this channel before
writing; session_resume/session_delete resolve refs against the
channel-scoped list.
- Shared limit helper extracted to tools.ListLimit.
**Telegram approval adapter (internal/channels/telegram/approval.go)**
- telegramApprover implements ApprovalRequester via the three-button
inline keyboard (Approve/Permanent/Deny). 2-min token expiry.
- Own map + mutex + "approval:" callback prefix — separate from the
dangerous-command system.
- Permanent approval uses exact-match (not prefix) lookup keyed by
action + resource description, so permanently approving "delete
session abc123" only covers that one session, not every call to
session_delete for 24h.
**Session lifecycle**
- /delete <session-id>: removes a session and its history. Refuses
prefixes shorter than 4 chars even when unique. Exact ID wins over
an ambiguous prefix. sessionTracker.forget sweeps every channel+
thread pointing at the deleted session.
- Extracted resolveSessionRef and shortSessionID for /resume and
/delete to share.
**Wiring**
- submitTurn builds the approver onto ctx before the LLM turn.
- chatGenerateOptions reads approver from ctx, passes it through
ToolSetOptions.
- SessionTools built per turn on the current message, combined with
taskTools into a single toolset.
Multi-identity daemon isolation:
- Split runIdentities into per-identity poll loops plus one shared
maintenance-and-drain loop, so reconcilePRs and storage cleanup run in
multi-identity mode and the global max_concurrency / per-repo
serialization holds across identities.
- Make the poll path identity-aware: each identity polls with its own
forge client, bot user, label, trigger, and ack reaction.
- Make allowConcurrentFor task-based and identity-aware via repoFor.
- Roll back storage mounts when container acquisition fails; log
Store.Update failures after branch set.
Container-mode RPC identity routing:
- forgerpc/worktreerpc gain identity-scoped subjects (SubjectFor,
RegisterFor) and identity-carrying clients.
- registerTaskRPCServers registers one server pair per identity (plus
root), so a container-mode task owned by a non-root identity is served
by its own forge client and worktree manager.
Gateway/Telegram:
- LLM-proposed session titles generated in the background after turns
(TitleGenerator + chatTitleGenerator), persisted by the gateway.
- /delete session command with prefix guard and tracker forget.
- Shared session display helpers (shortSessionID, sessionDisplayTitle,
resolveSessionRef).
- Reword docs and comments to drop vague "legacy" qualifiers where they no longer describe deprecation status
- Remove unused code: RecentEvents, MessageEvent.ToLegacy, index setState, stderr redaction tool, telegramOnlyCommands
- Delete or comment out tests covering removed or superseded behavior
Brings the chat surface to the Web UI: /api/chat with its own handlers
and tests, a self-contained ui/src/chat feature folder, and the shared
plumbing to reach the same router Telegram uses.
Includes a dangerous-command approval path (internal/gateway/dangerous.go,
internal/webui/dangerous.go, and the Telegram counterpart) so /rollback
and /stop are gated the same way on both surfaces, persona handling for
web sessions, and release-update formatting split out of the Telegram
service so both can render it.
web_parity_test.go pins that the two surfaces expose the same commands.
Authored in a parallel session; committed here so it lands with the rest
rather than sitting uncommitted while other work builds on top of it.
task check passes with it in place.
Screenshots used while building are gitignored (*.png) rather than
committed, matching the call made on the earlier design reference.
/branch copies the parent's history into the child. Because the store
honours an incoming MessageID -- which is what lets a read-modify-write
preserve identities -- the copies kept the parent's canonical IDs, so
two sessions claimed one identity and branch-point correlation was
ambiguous. Two branches from one parent made it three. It also broke
dedup in the child: an inherited message carried an ID derived from the
parent's session, so redelivering it appended (confirmed on both
backends: 3 messages became 4).
newMessageID's doc already claimed the opposite behaviour, and the
Message.MessageID doc said it was "ignored on write" -- which is exactly
the misreading that produced this. Both now describe what the code does,
including the obligation on a caller copying messages between sessions.
Also removes a doc comment describing msgID and its disambiguating
counter, deleted in 0814f7b, which had come to sit above
messageIDNamespace describing the opposite of the current design; and
corrects the Telegram comment still claiming a redelivered update
"updates the stored turn", the behaviour the immutability fix removed.
Refs archie-core-la1
go-telegram/bot retries a 429 immediately and only then grows its own
backoff, so a rate-limited daemon kept hammering getUpdates and stayed
limited. The pipeline error handler now reads Telegram's own RetryAfter
and waits it out, capped at a minute so a hostile value cannot park the
loop indefinitely, and cancellable so shutdown is not held up.
After the daemon was down ~9 hours it booted and answered every message
sent while it was gone -- 8 stale messages, 8 LLM turns, one burst.
Telegram holds undelivered updates for 24h and replays them on the first
getUpdates. launch()'s webhook branch already passes DropPendingUpdates;
the long-polling fallback had no equivalent, and polling is what this
deployment uses. deleteWebhook carries the flag even when no webhook is
registered -- the same call python-telegram-bot makes when it starts
polling -- and it doubles as a guard against a stale webhook, which
would otherwise fail getUpdates with 409.
Bounded by a 10s deadline so an unreachable Telegram cannot hold up
gateway startup.
The nestif extraction in 712388d pulled --provider handling out of
parseModelRest into a helper that took modelParts by pointer and
appended every surviving field to it. The caller's own loop appended
the same fields again, so every model name was doubled: /model gpt-5.6
asked the router for "gpt-5.6 gpt-5.6" and SetActiveModel rejected it
as unknown. The index loop becoming a range loop also stopped the
space-separated "--provider openai" value being consumed with its flag,
so it landed in the name too. Direct model switching over Telegram was
broken for everything but --refresh and provider-only switches.
splitProviderFlag is now the sole accumulator and returns the surviving
fields, so the shared-slice shape that allowed the double append cannot
recur (archie-core-0d6).
Found while auditing the remaining hunks of 6108fc5 (archie-core-ifi):
the naked return in the same function could not be regression-tested
until the function produced correct values. That audit is complete and
the other seven hunks are behaviour-neutral. Notably shell.go's
errorlint rewrite is not a repeat of the grep failure - it had no
companion exit-code condition to drop, and the ctx.Err() timeout check
already runs first - and find.go's errors.Is rewrite is safe because
errors.Is(err, nil) is false for any non-nil err.
Tests cover the return order of parseModelRest (two strings and three
bools, so a transposed pair compiles silently), the field count of the
model name, and shell's exit-code branch, which had no coverage at all.
The find fallback's habit of reporting a cancelled walk as "no matches
found" is filed as archie-core-ts9.
/stop cancelled the chat turn but could not touch a task the daemon was
executing, and /cancel refused outright:
case statusRunning:
return fmt.Errorf("task is actively running; cannot cancel
until it reaches a waiting state")
So the one case worth interrupting was the one case neither command
would touch.
The daemon now holds a cancel function for every executing task,
registered in process() -- the single point every dispatch path reaches
-- so a task is stoppable for its whole life, including the slow clone,
worktree prepare and image pull that is exactly when someone realises
they asked for the wrong thing.
- /cancel <id> interrupts the work first, then records the terminal
state. That order matters: writing the state while the task still runs
invites it to write its own afterwards and win. A task the store calls
running that nothing is executing (crashed daemon) still gets its state
recorded, which is what unsticks it.
- bare /stop now stops the reply, the queue behind it, and every running
task for the identity, reporting which of those it actually did.
Nobody reaching for the brake should have to look up a task ID first.
- StopRunning deliberately writes no terminal state: a stopped task
unwinds to parked, which is resumable. Forcing rejected would discard
work worth keeping.
- Without a runtime wired, cancelling a running task is still refused --
recording it as cancelled while it carries on would be a lie.
Container teardown detaches from the task context. Release runs on the
way out, and the main reason a task is on its way out is cancellation --
at which point the graceful ContainerStop failed instantly and the
container was left to the forced remove. Cleanup gets its own deadline.
process() gained two statements and crossed the funlen limit, so the
container setup block is extracted rather than suppressed.
/stop could not reach a running turn, and the reason was worse than a
missing cancel. go-telegram/bot runs one update worker that calls
handlers synchronously (defaultWorkers = 1, never overridden), and the
chat turn ran on it. While a turn streamed -- or sat in a shell tool --
the worker was blocked and /stop sat unread in the update channel until
the turn it was meant to kill had already finished.
Add gateway.Turns: a per-session serial lane that turns are submitted to
and that the event loop never waits on. The worker is free again in
microseconds, so /stop is serviced immediately; ordering comes from the
lane being serial rather than from the loop being blocked.
- Bare /stop cancels the running turn and clears the backlog behind it,
with no approval step. Approval guards destructive commands, and
gating the brake behind a confirmation defeats it. /stop <name> keeps
the existing process-termination behaviour and its approval.
- Cancellation propagates through the turn context, so it reaches the
model stream and any tool underneath -- a shell command built with
exec.CommandContext dies with the turn.
- Queues are unbounded and never shed. A dropped message is a silent
failure that surfaces later as an agent that ignored something.
- Turns are recovered per-lane: the channel's panic middleware wraps the
event-loop goroutine, which no longer runs the turn.
Also raise the chat step budget. It was hardcoded at 8 -- roughly one
read, one grep and two edits -- so turns hit the cap and stopped
mid-task with no indication they had been cut off. It is now
chat.max_steps, defaulting to 100. Interruption is /stop's job.
The gateway work added /whoami, /profile, /sessions, /resume and /agents as
executable commands; TestPublishedCommandsMatchExecutableCommandSurface
correctly caught that they were undiscoverable. Two workers touching a
coupled surface -- the test is the thing that noticed.
/new, /reset, /topic, /retry, /undo, /title, /branch, /fork, /compress,
and /compact are executable via the gateway router but were missing
from gatewayCommandSpecs, making them undiscoverable in Telegram's
command menu and /help output.
- bodyclose: add deferred resp.Body.Close() in storage_test.go Docker test handler
- contextcheck: thread context through email handleSMTP/processMessage; replace
context.Background() with context.WithoutCancel(ctx) in telegram, webhook, webui
shutdown paths and container pool cleanup; add ctx parameter to Pool.Release
- containedctx: replace ctx context.Context struct field with stopCh chan struct{}
in StdioTransport and SSETransport; thread lifecycle context through goroutine
chains instead of storing it in structs
- fatcontext: lift context deadline derivation out of PrefetchContext loop
Add gateway/runtime-owned commands that remain usable when LLM credits
are exhausted:
- /start — acknowledgement without consuming a reply
- /new [name] (alias: /reset) — create a fresh session
- /topic [off|help|<session-id>] — session switching for threaded chats
- /retry — remove last assistant message, replay last user turn
- /undo [N] — remove last N messages from history (default: 1)
- /title [name] — set or display session title
- /branch [name] (alias: /fork) — create child session with inherited history
- /compress [here N|focus topic|--preview|--dry-run] (alias: /compact)
Design:
- sessionTracker maps channel+thread to active session, auto-creating
implicit sessions with the legacy sessionKey format for backward
compatibility with existing persisted data
- /new and /branch create explicit UUID-backed sessions
- SessionStore extended with DeleteRecentMessages, MessageCount, and
SaveMessages for undo/retry/branch/compress operations
- SessionContext extended with Title, ParentSessionID, BranchName
- All commands report 'not configured' when session store isn't wired
Session resolution is wired into the LLM responder so /new, /branch,
and /topic switches affect subsequent chat turns.
Implement session-scoped goal/subgoal/steer/queue state management
with thread-safe access and prompt injection for standing goals.
Commands:
- /steer <text> — queue a message to inject after the next tool call
- /goal [text|draft|show|pause|resume|clear|status|wait|unwait]
- /subgoal [text|remove N|clear]
- /queue (aliases /q, /bg, /btw, /background) [text|remove N|clear]
Architecture:
- SessionState in internal/gateway holds per-session goals, subgoals,
steer queue, wait flag, and background queue
- Goals are injected into the system prompt via BuildGoalPrompt()
- Steer messages are consumed between agent tool-calling steps via a
MaxSteps=1 loop in the daemon's respond closure
- Commands route through the existing gateway.Router with not-configured
guards when State is nil
Closes: archie-core-alm.1
Extend /model command to support:
- /model [model] (direct switch)
- --provider <name> (provider switch)
- --global (persist scope, with ModelPersist callback)
- --session (explicit session scope, default)
- --refresh (show current active model)
Add /personality command:
- /personality lists available personalities with inline keyboard
- /personality <name> switches directly
- Callback handling for inline keyboard selection
Router changes:
- Add Personas field (wired to PersonaRegistry in main.go)
- Add ModelPersist callback for --global persistence
- Route /personality as a local command
- Extend handleModel to accept full rest text with flag parsing
Gateway-local commands that report from Archie-owned state:
- /whoami: reports the Archie identity the user is talking to
- /profile: shows active Archie profile (identity, provider, model, available models)
- /sessions: lists recent sessions with last-active timestamps
- /resume [id]: resumes a session by ID or prefix match
- /agents: lists active agents and running/waiting tasks
Adds SessionLister and AgentReader interfaces to Router, wired from the
daemon composition root. SessionStore already satisfies SessionLister.
AgentReader is a thin adapter over store.Task list filtered to active
states (running, waiting_human, parked).
All commands are published in Telegram's menu and /help, and respond
with 'not configured' when their backing interface is nil.
Closes archie-core-alm.2
Hide /approve from Telegram's command menu, help text, and discoverability.
Telegraph users can no longer learn about task approvals through the bot.
/approve remains executable in the gateway router for email and webhook
channels; non-Telegram workflow callers are unaffected. Daemon polling of
forged waiting_human tasks and the Requeue(→implement) handoff are untouched.
97 -> 85. Not one of the eleven nilerr findings was a bug.
I called these "real bug candidates" when I planned the phases. I was
wrong, and the reason is worth recording: every site is the same
deliberate idiom, and a linter cannot see intent.
- inprocess.go (5) and gateway.go (2) are handlers that report a
problem as TEXT to their consumer. A tool handler returning a Go
error aborts the tool call; returning "rejected: ..." lets the model
read the feedback and retry with valid arguments. A chat handler
returning a Go error surfaces as a gateway fault instead of a usage
hint to the human.
- worker.go carries runErr in the response envelope's Error field.
Returning it as a Go error too would double-report and abort the
NATS reply.
- email.go returns nil when Accept fails because context cancellation
closed the listener. That is a clean shutdown.
- codesearch.go skips an unreadable entry and continues the walk. A
best-effort index must not abort on one bad file.
- skill.go is documented in the function comment already: a
non-directory skills path means "no skills", not a fatal error.
Annotated each with //nolint:nilerr and the specific reason rather than
excluding the linter by path, so nilerr stays live everywhere else and
the next reader wondering "why nil here?" gets an answer at the site.
This is the phase I insisted on keeping single-session and out of the
subagent fan-out. A worker told to "fix nilerr" would have converted all
eleven to error returns: malformed tool arguments would hard-fail the
agent loop instead of being correctable, chat typos would read as
gateway faults, and shutdown would log spurious errors. Every test would
still have passed.
130 -> 96. forcetypeassert, predeclared and usetesting are now at zero.
Fanned out across four isolated worktrees partitioned by package so no
two workers could touch the same file: internal/tools (20 assertions),
internal/webui, internal/memory + internal/ratelimit, internal/secret.
Four stragglers outside those scopes fixed inline (internal/pairing,
internal/forge, internal/channels/telegram, internal/tools/mcp).
Unchecked type assertions became comma-ok with a t.Fatalf naming the
actual type. One site in a goroutine uses t.Errorf + return instead,
since t.Fatalf from a non-test goroutine is undefined behaviour.
predeclared renames chose names from meaning rather than adding a
suffix: max -> maxQueryLen, maxMatchLen, maxRequests, alphabetLen, and
nextBackoff's max -> ceiling. `real` -> `base` in the webui tests.
os.Setenv became t.Setenv, which restores on cleanup, so the paired
deferred Unsetenv calls went with it.
Note for future fan-outs: the agent worktrees branched from origin/main
(a7388a0), not from this branch, so every diff arrived based on a tree
without the phase 1 autofix. They were applied with `git apply --3way`
and the phase 1 changes to internal/memory/manager.go were verified to
survive — in particular the repaired `ctx, cancel =` assignment. A plain
`git apply` would have silently reverted it.