tickmarkr 1.54.0 → 1.55.0

package/README.md

@@ -173,8 +173,56 @@ More levers, all optional and absent-safe (absent config keeps default behavior)
   the routed model's window
 - `routing.sla` — per-shape latency expectations, surfaced as advisory plan lints against the
   learned performance profile
-- `run --quality` — raise every capability floor one band for that run (it never lowers one),
-  disable exploration, and record the original and raised floor in routing provenance
+- `run --quality` — compatibility alias for `run --mode partner-led` for that run; it disables
+  exploration and records the selected mode in routing provenance
+
+## Steering: routing modes, reviewer preferences, and reruns
+
+### Routing modes
+
+`routing.mode` is a preset that compiles into floor assignments at config load time. The router never sees the mode itself; it receives resolved floors only.
+
+Three routing modes are available:
+
+- **`risk-based`** (default): byte-identical to pre-v1.51 routing. Absent `mode` key resolves as risk-based.
+- **`partner-led`**: resolves every non-overridden shape to a `frontier` floor and disables exploration — use when quality is paramount and cost is secondary.
+- **`staff-led`**: lowers each mode default by one tier (e.g., `implement` and `refactor` become `cheap` instead of `mid`) while keeping the preset floor for the integrity set (`plan`, `spec`, `migration`, `ui`) at `frontier`.
+
+Explicit `routing.floors` entries beat mode-preset deltas and are linted during plan if they shadow the mode's delta; an explicit integrity floor below `frontier` is also linted. The mode is compiled once at config load and never consulted during routing.
+
+### Review preferences
+
+`review.prefer` is an ordered list of reviewer seats for the cross-vendor code-review gate. Entries are matched by diversity (never the same vendor or model as the original worker), and routing reorders the available channels only — it does not admit unauthed or denied channels.
+
+```yaml
+review:
+  prefer: [codex, kimi]           # bare adapter: inherits model from the routed channel
+  prefer: [codex:gpt-5.6-sol, kimi:kimi-code/k3]  # adapter:model explicit
+  prefer: [codex, kimi:kimi-code/k3]  # mixed: bare and explicit
+```
+
+**Grammar**: review prefer entries may name a bare adapter (inheriting the model from the current channel) or an explicit `adapter:model` pair. Bare adapters rank every diversity-eligible channel for that adapter; explicit pairs rank one diversity-eligible channel.
+
+### Consult preferences
+
+`consult.prefer` is a ranked failover list of seats for escalations on deadlock or gate stalls. Unlike review, a consult seat has no channel to inherit a model from, so entries **must be explicit `adapter:model` pairs**.
+
+```yaml
+consult:
+  adapter: claude-code
+  model: fable
+  prefer: [codex:gpt-5.6-sol, kimi:kimi-code/k3]   # adapter:model ONLY
+```
+
+The daemon walks the preference list to the first live adapter, then the pinned `consult.adapter:model` pair as the final fallback. Failed or unparseable verdicts fall to the next entry.
+
+**Grammar**: consult prefer entries require `adapter:model` form — a bare adapter name is invalid and fails config load. Every entry must declare both the adapter and the model because a consult seat runs independently with no channel context.
+
+### Rerun control: --supersedes
+
+`tickmarkr run --supersedes <prior-runId>` marks the current run as a rerun of a prior engagement. The current task graph is used for the rerun; compile it fresh first if the spec changed. The prior runId is recorded in the new journal, and the prior journal records the successor, for audit trails and change attribution.
+
+Use this when you modify the spec or worker logic and want to mark an intentional rerun while preserving the relationship in both run journals.
 
 ## Model scoping and auth detection
 

package/dist/adapters/registry.js

@@ -77,6 +77,19 @@ const MODEL_PROBE_PROMPT = "Reply with exactly OK and nothing else.";
 const MODEL_PROBE_TIMEOUT_MS = 60000;
 // Auth-words only match when tied to a failure word; bare "auth"/"OAuth"/"authored" never fail (v1.27 T2).
 const AUTH_FAILURE_RE = /\b4\d\d\b|\bauth(?:entication|orization)?\s+(?:error|failed|failure|denied)|unauthori[sz]ed|forbidden|access denied|credit(?:s)?\s+(?:exhausted|error|denied)/i;
+const PROBE_REASON_CAP = 240;
+// v1.55 T3: the tail must open at a word boundary — a mid-word slice ("odel 'grok-…'") reads as
+// corruption in operator-facing diagnostics. Input is already space-normalized, so " " is the only
+// boundary. A tail that is one unbroken token keeps its mid-word cut rather than storing nothing.
+function reasonTail(output) {
+    if (output.length <= PROBE_REASON_CAP)
+        return output;
+    const tail = output.slice(-PROBE_REASON_CAP);
+    if (output[output.length - PROBE_REASON_CAP - 1] === " ")
+        return tail;
+    const sp = tail.indexOf(" ");
+    return sp === -1 ? tail : tail.slice(sp + 1);
+}
 function probeFailure(code, stdout, stderr, timedOut, timeoutMs = MODEL_PROBE_TIMEOUT_MS) {
     // SIGKILL-timeout is not exit-1: report the budget, never the masked kill code (v1.27 T1).
     if (timedOut)
@@ -85,7 +98,7 @@ function probeFailure(code, stdout, stderr, timedOut, timeoutMs = MODEL_PROBE_TI
     // OBS-72: TAIL, not head — the error lands at the END of CLI output; a head slice stores only the
     // startup banner and hid the real "Not inside a trusted directory" failure for a day.
     return code !== 0 || QUOTA_RE.test(output) || AUTH_FAILURE_RE.test(output)
-        ? output.slice(-240) || `probe exited ${code}`
+        ? reasonTail(output) || `probe exited ${code}`
         : undefined;
 }
 export const pendingAutoPreferKey = Symbol.for("tickmarkr.pendingAutoPrefer");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tickmarkr",
-  "version": "1.54.0",
+  "version": "1.55.0",
   "description": "Spec in, verified work out.",
   "type": "module",
   "license": "MIT",

package/skills/tickmarkr-loop/SKILL.md

@@ -50,6 +50,10 @@ Use one of:
 
 After sending, **confirm delivery** by reading the target pane and verifying the message landed (input empty, agent status `working`, or notification acknowledged). Never report "briefed" or "relayed" without read-back confirmation.
 
+## Dedicated consultant tab rule
+
+When spawning consultants (agents gathering synthesis input for decisions like SCOPER analysis or architectural reviews), create them in a DEDICATED tab separate from the ORCHESTRATOR tab. This ensures that when the orchestrator stands down, the consultant panes persist and their assessments remain available for review and reference.
+
 ## Stand-down (mission end and retirement)
 
 - **Orchestrator, on terminal state** (green, failed, or parked), after the record commit and operator notification: stop every monitor and background task you started, print one final stand-down line, and leave NOTHING queued in your input box. A finished session with an armed watcher or pre-filled input is a loaded gun — a retired v1.40 orchestrator sat idle with "merge … tag, publish" unsent in its input; one stray Enter would have shipped a duplicate release.