ftown-bridge 0.19.10 → 0.19.11

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ftown-bridge",
-  "version": "0.19.10",
+  "version": "0.19.11",
   "description": "CLI bridge for ftown — generic PTY-over-Centrifugo relay",
   "type": "module",
   "main": "dist/index.js",

package/skills/factory/SKILL.md

@@ -34,15 +34,21 @@ already initialized (point them at `status`/`up`/`down` instead) — do not over
    `uv` is on PATH.
 2. **Install fticket.** `uv tool install fticket` (or `uv tool upgrade fticket` if already
    installed). Verify with `fts --help`.
-3. **Copy the template.**
+3. **Resolve routing and copy the template.** If the deployment prompt names the
+   initiating harness/model, use those values. Otherwise, when `FTOWN_SESSION_ID` is
+   set, run `~/.ftown/ftown-sessions get "$FTOWN_SESSION_ID"` and read this session's
+   `shellType` and optional `model`. If neither source identifies the harness, ask the
+   user instead of assuming one.
    ```bash
    cp -r ~/.ftown/skills/factory/factory-template "$REPO/factory"
    ```
-   `factory/bin` moves in as-is (no edits). Ask the user for: project name (default: the
-   repo directory's basename), an operator ftown session id (optional — default `-`, no
-   operator mail), any stage routing tweaks (harness/model/max_workers per stage), and
-   the shared resources workers must lease (default `staging`, exclusive-only).
-   Edit `factory/factory.yaml` with their answers.
+   `factory/bin` moves in as-is (no edits). Set every stage plus `triage` and `digest` to
+   the initiating harness/model. When the initiator has no explicit model, remove each
+   template `model` value so the harness default is used. Only retain or add a different
+   route when the user explicitly requested that override. Also apply the project name
+   (default: the repo directory's basename), operator ftown session id (optional —
+   default `-`, no operator mail), and shared resources (default `staging`,
+   exclusive-only). Edit `factory/factory.yaml` with these values.
 4. **Gitignore + state dirs.** Append `.ffactory/` to `$REPO/.gitignore` (create the file
    if absent). `mkdir -p "$REPO/.ffactory/tickets"`.
 5. **Init the ticket db**, reading stage order and limits back out of the edited

package/skills/factory/factory-template/bin/dispatch.py

@@ -272,7 +272,11 @@ def capacity_spawn(
     db_path: Path,
     dry_run: bool,
 ) -> tuple[list[int], int, int, int]:
-    """Claim + spawn up to capacity, stages in yaml order.
+    """Claim + spawn up to capacity, later stages first (pipeline drain).
+
+    Reversing yaml stage order prefers tickets already moving through the pipeline over
+    admitting new work at the first stage. With max_sessions=1, one ticket therefore
+    runs end-to-end before the dispatcher starts the next one.
 
     Returns (spawned_ids, rejected_input, expiry_capped, active_total). Guards run just
     after the claim (fts.claim picks the ticket, so preflight can't run earlier); a guard
@@ -297,7 +301,7 @@ def capacity_spawn(
         except FTSError as exc:
             print(f"dead_letter failed for ticket {ticket_id}: {exc}", file=sys.stderr)
 
-    for stage in cfg.stages:
+    for stage in reversed(cfg.stages):
         free = min(stage.max_workers - active_of(stage.name), budget)
         if free <= 0:
             continue

package/skills/factory/factory-template/factory.yaml

@@ -5,7 +5,7 @@ project: my-project            # used to namespace ftown loops and worker names
 operator_session: "-"          # ftown session id for digests/worker mail; "-" = none
 
 limits:
-  max_sessions: 6              # global concurrent worker cap for this factory
+  max_sessions: 6              # global cap; set 1 for strict one-ticket end-to-end flow
   bounce_limit: 3              # rejections before auto dead-letter (fts config)
   claim_ttl_ms: 1800000        # default claim TTL (30m); stages may override
   max_ticket_age_h: 48         # dispatcher dead-letters tickets older than this
@@ -17,6 +17,7 @@ resources:
   - name: staging
     policy: exclusive_only     # both | exclusive_only | shared_only
 
+# Dispatcher claims later stages first, draining work before admitting new tickets.
 # Harness per stage: claude | codex | cursor | zai | kimi | deepseek | fireworks
 # next: where advance() sends the ticket on success (- = terminal)
 # bounce: where reject() sends it (- = rejection not allowed)