diff --git a/plans/tcl-sx-completion.md b/plans/tcl-sx-completion.md
new file mode 100644
index 00000000..39f633bd
--- /dev/null
+++ b/plans/tcl-sx-completion.md
@@ -0,0 +1,146 @@
+# Tcl-on-SX completion plan — SX capabilities first
+
+Tcl phases 1–6 are complete (329/329 tests). This plan covers the remaining
+limitations, ordered by the SX work needed to enable them.
+
+## Key audit findings
+
+Several apparent gaps are already solved in SX:
+
+- **Floats** — SX parses `3.14` natively; `(+ 1.5 2.5) → 4.0`; `str` formats
+  with `%g` (compact, no trailing zeros). `floor`/`ceil`/`round`/`truncate`
+  all exist in `spec/primitives.sx`.
+- **Regex** — `regexp-match`, `regexp-match-all`, `regexp-replace`,
+  `regexp-replace-all`, `regexp-split` are registered OCaml primitives using
+  `Re.Pcre` (`hosts/ocaml/lib/sx_primitives.ml`).
+- **`call/cc` multi-shot** — works. `set!` on closed-over vars works. Fibers
+  are implementable as a pure SX library.
+- **`perform` user-accessible** — `(perform :foo 42)` from user code suspends
+  the evaluator and emits an IO request. The algebraic effects model is
+  already half-built.
+- **No `file-read`/`clock-seconds`** — not yet registered as OCaml primitives.
+  Only string ports exist. Would need small OCaml additions.
+- **No `env-as-value`** — environments are internal OCaml values, not
+  inspectable from SX user code.
+
+---
+
+## Phase 1 — Zero-cost wins (no SX changes, only `lib/tcl/`)
+
+Everything here is pure Tcl implementation work.
+
+| Work | Effort | Unlocks in Tcl |
+|---|---|---|
+| Float in `expr` — detect `.` in number tokens, route through float ops instead of `parse-int` | half day | `expr {3.14 * 2}`, `expr {sqrt(2.0)}`, float comparisons |
+| `regexp pattern str` and `regsub pattern str repl` wrapping existing SX primitives | few hours | pattern matching, text processing |
+| `apply {args body} ?arg…?` — anonymous proc call | 1 hour | higher-order functions, `lmap` idiom |
+| `array get/set/names/size/exists/unset` commands | half day | array variables (tokenizer already parses `$arr(key)`) |
+
+**Total: ~2 days. Zero SX changes.**
+
+---
+
+## Phase 2 — `lib/fiber.sx` (pure SX library, no OCaml)
+
+`call/cc` is multi-shot and `set!` on closed-over vars both work. Fibers are
+implementable as a pure SX library using symmetric continuation swapping:
+
+```scheme
+; lib/fiber.sx — canonical fiber primitive for all hosted languages
+(define make-fiber
+  (fn (thunk)
+    (define slot-k nil)
+    (define slot-caller nil)
+    (define slot-done false)
+    (fn (resume-val)
+      (call/cc (fn (caller-k)
+        (set! slot-caller caller-k)
+        (if (nil? slot-k)
+          (begin (thunk resume-val) (set! slot-done true) (caller-k nil))
+          (slot-k resume-val)))))))
+
+(define fiber-yield
+  (fn (val)
+    (call/cc (fn (k)
+      (set! slot-k k)
+      (slot-caller val)))))
+```
+
+Each coroutine becomes a fiber. `yield` swaps to the caller; calling the
+coroutine name swaps back. True suspension, not eager pre-execution.
+
+**Broader value:** Ruby fibers, Python generators, Lua coroutines, async event
+loops, cooperative schedulers all sit on top of the same library.
+
+**Alternatively:** `perform` is user-accessible. A Tcl scheduler living outside
+the SX evaluator (the OCaml host or an SX event loop) could catch
+`(perform :fiber-yield val)` and dispatch it — the algebraic effects model,
+already half-built.
+
+**Total: 2–3 days. Produces `lib/fiber.sx` as a lasting SX contribution.**
+Tcl coroutines then rewrite using `make-fiber` for true suspension.
+
+---
+
+## Phase 3 — Small OCaml additions (`sx_primitives.ml`)
+
+Each is ~10–20 lines of OCaml. All are useful across the whole platform, not
+just Tcl.
+
+| Primitive | OCaml effort | Unlocks |
+|---|---|---|
+| `(file-read path)` → string | tiny | Tcl `open`/`read`, SX scripts reading files |
+| `(file-write path str)` → nil | tiny | Tcl `open`/`puts` to files |
+| `(file-exists? path)` → bool | tiny | Tcl `file exists` |
+| `(file-glob pattern)` → list | small | Tcl `glob` |
+| `(clock-seconds)` → int | tiny | Tcl `clock seconds` |
+| `(clock-format n fmt)` → string | small (wraps `strftime`) | Tcl `clock format` |
+
+**Total: 1 day. One focused afternoon of OCaml.**
+
+---
+
+## Phase 4 — Optional: env-as-value (architectural)
+
+`uplevel`/`upvar` required an explicit frame stack because SX environments
+aren't inspectable from user code. Adding:
+
+```scheme
+(current-env)            ; → env value
+(eval-in-env env expr)   ; → result
+(env-lookup env key)     ; → value or nil
+(env-extend env key val) ; → new env (non-mutating)
+```
+
+...would let `uplevel N` be literally "look up env N levels up, eval in it."
+The Tcl frame stack (hundreds of lines) collapses to ~10 lines.
+
+Also benefits: metacircular evaluators, REPL tooling, live debugging (inspect
+any scope), the sx_docs server's eval endpoint.
+
+More invasive — touches `sx_types.ml` and `sx_server.ml` — but a meaningful
+architectural improvement worth doing when the moment is right.
+
+**Total: 2–3 days. High architectural value, not urgent.**
+
+---
+
+## Suggested order
+
+1. **Phase 1** — immediate Tcl wins, zero risk, proves the approach
+2. **Phase 2** (`lib/fiber.sx`) — the interesting SX work, benefits all hosted languages
+3. **Phase 3** (OCaml primitives) — quick practical completions
+4. **Phase 4** — architectural cleanup when it's worth the invasiveness
+
+Phases 1+2+3 ≈ one focused week. Tcl is genuinely complete, and `lib/fiber.sx`
+becomes a lasting SX contribution used by every future hosted language.
+
+---
+
+## What stays out of scope
+
+- `package require` of binary loadables
+- Full `clock format` locale support  
+- Tk / GUI
+- Threads (mapped to coroutines only, as planned)
+- Full POSIX file I/O (seek/tell/async) — stubs are fine