Backlog

Things decided worth doing and not done yet, so they are not forgotten. Each entry says where the idea comes from and what the work is. Move an entry to CHANGELOG.md when it ships; add a line here whenever a review or a phase defers something.

From the FastFlow review (2026-09-06, decisions/0002-fastflow-review.md)

Item	FastFlow source	The work
Divide and conquer pattern	dc.hpp, tests/test_dc.cpp	tq.divide_and_conquer(divide, combine, base, is_base, workers): a farm with feedback whose collector counts pending subtasks per parent and combines when they are all back. Fibonacci and mergesort as tests.
Parallel-for, reduce and map helpers	parallel_for.hpp, map.hpp	tq.map(fn, items, workers=8, ordered=True) returning a list; tq.parallel_for(n, body, workers, grain); tq.parallel_reduce(...). Thin helpers over scatter and on-demand farms, for people who want a result, not a graph. A session keeps the threads between calls in a loop.
Per-input end hook	ff_node::eosnotify(id)	on_input_end(self, ctx, source) on class nodes, called when one input closes and allowed to send; today only raw nodes see it through recv(source=i) returning None. Joins emit partial results earlier.
Byte cap on network batches	batchByteSize in the DFF config	A size limit next to the item limit on TCP batches, per group in the deploy file, so big items do not build huge frames and small ones still batch.
Thread pinning	ff_mapThreadToCpu, FF_MAPPING_STRING	tq.run(graph, pin="0,2,4") or pin=True using os.sched_setaffinity for node threads and child processes on Linux. Benchmarks on 3.14t first, to see whether it pays.
Resize a farm while it runs	run_then_freeze, thaw(nw), the manager channel	session.set_workers(farm, n): the emitter stops feeding workers beyond n; they idle until raised again. Adaptivity and energy.
MPI transport	ff_dsenderMPI, MTCL	An mpi4py transport behind the same frames as TCP, launched through mpirun from the deploy file. Only if a cluster user asks.
Zero-copy check	June 2026 commit on owner-aware serialization	Confirm the pipe and TCP paths pass pickle protocol 5 out-of-band buffers for numpy arrays, and add a benchmark row.

From earlier phases

Item	Where it came up	The work
Nested feedback loops	Feedback.expand refuses feedback(feedback(...)) and a node in two loops	Loop tokens per loop with a stack of loop memberships; the head of an inner loop closes before the outer one can.
Subinterpreter runtime	DESIGN.md runtime table, Phase 6	One subinterpreter per node (PEP 734, 3.14) behind runtime="interpreters", channels through interpreter queues.
Farm of blocks with ordered or gather collection	1.1 nested blocks	Tags would have to travel through every node of the copy; today ordered and gather farms keep plain node workers.
Distributed runtime on Windows	tests/test_net.py runs on loopback only	Verify sockets and the launcher on Windows and macOS; the CI matrix covers threads and processes there.

Tolquane Web

Deferred while the first release was built (web.md, "Later, not in the first release", and the I3 hardening pass), and what 1.3 left for later.

Item	Where it came up	The work
A workspace per user, and permissions per flow	web.md, one server one workspace; 1.3 U	Accounts with roles shipped in 1.3, but everybody who signs in shares the one workspace and may run anything in it, as the user who started the server. A workspace per user, or permissions per flow, is still a change of model rather than a setting.
HTTPS	I3 security review; 1.3 U	A --tls-cert and --tls-key pair, or a documented reverse proxy. Today a token on a shared machine travels in clear, and since 1.3 so do passwords and session tokens; the guide says to use an SSH tunnel instead.
Breakpoints that pause a node	web.md	Hold an item at a node until the user releases it, on the sync runtime first; needs a control channel into the run's child process, which the events stream does not have.
Deploy editor	web.md	Assign cards to groups on the canvas, write deploy.toml, launch every group with tolquane launch and watch them all in one page.
Gallery of templates	web.md	More than empty and hello: the examples, and the flows the AI builder wrote, offered as starting points from the New flow dialog.
Sharing a flow as a link	web.md	A read-only page for one flow and one run, addressable and safe to send; needs a signed link that carries its own token, or a read-only role beside admin and member.
Per-user rate limiting on the AI panel	I3 hardening	The chat spends the owner's key and has no limit of its own beyond the provider's; a request counter per hour would make an accident cheaper.
Events on disk for a long run	I3 hardening	Events are kept in memory and bounded (MAX_EVENTS), so a very long run loses its middle. Writing them to a file under .tolquane-web/ would let the drawer scroll the whole run back.
Retries and notifications that survive a restart	1.3 B3	A pending retry is a threading.Timer and a delivery is a daemon thread, so a server that stops between two attempts drops the chain and a message in flight is lost. A queue in the store, drained at startup, would make both survive.
Notifications beyond a webhook and email	1.3 B3	Slack, Teams and the like go through the webhook today, in Tolquane's own JSON shape; per-target templates, or first-class integrations, if anybody asks.

Platforms

Item	Where it came up	The work
Cancel on Windows	tolquane run --events and the run supervisor	Windows has no SIGTERM: a cancel kills the child, which ends as failed rather than cancelled. A control pipe or a file flag would let the child stop cleanly.
The network runtime on macOS	tests/test_net.py is expected-to-fail on the macOS runner	Loopback connections between groups on the macOS runner sometimes drop, or a sender never comes back within the budget after a drop; every test passes on Linux and Windows. Reproduce on a Mac, then remove the mark.