@mrciphersmith/keryx 0.3.68 → 0.3.75

package/README.md

@@ -180,6 +180,7 @@ alike, whichever agent runtime happens to be open.
 | Judge readiness | Normalized health reports and a quality gate over lint, types, tests, coverage, complexity |
 | Coordinate work | Versioned task flows, managed review packages, generated agent skills |
 | Keep agents inside boundaries | Deterministic secret / PII / prompt-injection scanning, redaction, policy gate, OS sandbox |
+| Take work in from GitHub | Tickets assigned to you, review requests, failed CI and PR comments arrive as Telegram cards with buttons, read-only toward GitHub ([guide](docs/docs/guides/work-intake.md)) |
 | Run an agent at all | A first-party harness on top of all of the above: durable sessions, allow/ask/deny policy, child agents, evidence-gated completion |
 
 ## A typical agent workflow
@@ -709,16 +710,23 @@ Grouped by what you are trying to do, not by internal module layout.
   order, your choice, the time you took and, after a deviation, an optional
   reason (asked once, and the tool result waits for it; an empty answer releases
   the wait; add or change it later with `/decisions reason`).
-  One question in three is asked blind (no "recommended" mark, random order,
-  revealed after the answer, changeable with `/decisions change`), never for
+  A question with a recommendation is asked in one of four ways: ordinary (arm A, the
+  recommended option preselected), the mark with nothing preselected (B), shuffled with the
+  mark (C), or blind (D: no "recommended" mark in the flag or the words, random order,
+  revealed after the answer, changeable with `/decisions change`). The arm is a seeded
+  function of a per-repository salt; the weights are 0.4, 0.2, 0.2, 0.2. Never blind for
   release, ship, publish, deploy, delete or push, nor for merge, drop, remove or
   force next to main, production, a branch or a table, or with no real object
   ("merge it"); agents set `--action` (or `irreversible: true` in `ask_user`) for
   the irreversible ones, and the report counts how often blind was refused.
-  `keryx decisions report` prints the match share by mode and stage and the
-  deviations, with no model; `keryx decisions open|answer|reason` lets any agent
+  `keryx decisions report` prints the match share by mode, arm, channel and stage and
+  the deviations, with no model (the pre-arm records are shown apart as legacy);
+  `keryx decisions rate` records whether a recommendation was good (yours, or a
+  model's self-assessment in a clean context) and `keryx decisions export` writes
+  the structure of each decision with no question or option text;
+  `keryx decisions open|answer|reason` lets any agent
   or chat bridge drive it. In the TUI, `/decisions` and a sidebar row show the
-  same report. See the
+  same report. The guide says what the journal does not measure. See the
   [Recommendation journal guide](docs/docs/guides/recommendation-journal.md).
 - **security** — deterministic secrets / PII / prompt-injection / egress
   scanning, redaction, and a policy gate at agent write seams, with a committed
@@ -742,7 +750,11 @@ Grouped by what you are trying to do, not by internal module layout.
   a working session keeps, plus an external hand onto it: the `slate.open` /
   `slate.writeSeed` / `slate.close` MCP tools let any MCP-connected harness open
   one scoped to its own session id and close it into the same review pipeline.
-  Local stdio only; there is no CLI verb. See the
+  Local stdio only; there is no CLI verb. In keryx's own sessions the slate is
+  also the model's working memory: a harness-written Trail of every tool call,
+  model-written Notes (`slate_note`), and read-only recall (`slate_trail`,
+  `recall_step`, `history_search`) replace re-sending old rounds; `/trail` and
+  `/notes` show you the same Trail and Notes in the shell. See the
   [Slate guide](docs/docs/guides/slate.md).
 
 **Run agents inside boundaries**
@@ -1040,14 +1052,19 @@ queued lines will not run. From the topic you can also run a fixed list of slash
 commands (`/status`, `/model` and `/connect` as buttons, `/new`, `/clear`, `/resume`
 in the same topic); `/mode trust`, `/plan off`, `/delegate` and `/external*` run only
 after a Yes press, and commands such as `/mcp`, `/guard`, `/route`, `/provider`,
-`/schedule`, `/rewind`, `/conform` and `/ci` stay local. Each message shows its state as a reaction and the topic shows typing
+`/schedule`, `/rewind`, `/conform` and `/ci` stay local. A topic is deleted when the
+shell exits, so `/history [N]` (1 to 20, 10 by default) posts the session's last messages
+back into the topic, yours and the agent's final answers, one message each; a resumed
+session whose topic was just created gets the last 10 by itself, once (the first time remote control is turned on in that shell); a turn that looks like it holds a secret is replaced by a placeholder. Each message shows its state as a reaction and the topic shows typing
 while a turn runs; see
 [commands from the topic](docs/docs/guides/drive-keryx-remotely.md#commands-from-the-topic).
 Without a connection `serve` starts as before, with remote control off, and only
 offers the local `/channels` routes. Replies are rendered in one of four modes,
 `remote.rendering` = `auto` (the default), `rich`, `html` or `plain`, set from the
 `/settings` row **Telegram rendering**: a table goes out as a native Telegram rich
-message in `auto` and `rich` (an aligned `<pre>` block in HTML), ordered and nested
+message in `auto` and `rich`, as wide as its content and scrolling sideways with no
+cell shortened (an aligned `<pre>` block in HTML; more than 20 columns stays ordinary
+text), ordered and nested
 lists, task items and rules keep their shape, and a refused rich message falls back
 once to HTML and once to plain text, so a reply is never dropped. `keryx remote
 format-sample` prints the sample in every mode with no network, and `/channels`