kres-repl: halt on N consecutive task errors and surface them to the …

.githooks/commit-msg

@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+# Commit-msg hook for the kres repo.
+#
+# Rejects any commit whose message has a prose line longer than 100
+# characters. Lines starting with `#` are git's editor template
+# comments and are stripped before the message is recorded — they're
+# skipped here so the hook matches what git actually stores.
+#
+# To install:   git config core.hooksPath .githooks
+
+set -euo pipefail
+
+msg_file="$1"
+limit=100
+
+bad=$(awk -v limit="$limit" '
+    /^#/ { next }
+    { if (length($0) > limit) printf "  line %d (%d chars): %s\n", NR, length($0), $0 }
+' "$msg_file")
+
+if [[ -n "$bad" ]]; then
+    cat >&2 <<EOF
+commit-msg: rejecting message — prose line(s) over ${limit} chars:
+${bad}
+
+Re-wrap the offending line(s) and retry. Use --no-verify only if you
+must bypass.
+EOF
+    exit 1
+fi

CLAUDE.md

@@ -54,6 +54,34 @@ User prompt → Task created → Task thread starts
 - Auto-progress checks goal after each completed task for early exit
 - Deferred items (identified but not started when goal met) saved via `/followup`
 
+### Plan + Session Persistence
+- `kres_core::Plan` holds the planner's decomposition: `prompt`, `goal`,
+  `mode`, and `steps` (each with `id`, `title`, `status`, `todo_ids`
+  linking to `TodoItem` rows). Lives on `TaskManager` as
+  `Option<Plan>`; `sync_plan_from_todo` rolls up step status from
+  linked todo statuses.
+- `kres_core::SessionState` (`<results>/session.json`) is the
+  resumable snapshot: plan + todo list + deferred list +
+  `completed_run_count` + last prompt. Written atomically (tmp +
+  fsync + rename) from the reaper tick and the various drain
+  paths.
+- Resume: `kres --results <dir> --resume` loads the snapshot,
+  flips every `InProgress` todo/plan step back to `Pending` (its
+  prior executor is gone), and seeds the manager + deferred list
+  before the REPL starts. Without `--resume`, any existing
+  session.json is left untouched on disk and the REPL starts
+  clean; a note in the startup banner points at the file so the
+  operator knows the state is recoverable.
+- InProgress drains: ctrl-c, the `--turns N` cap, goal-met, and
+  `--turns 0` follow-stagnation all call
+  `TaskManager::reset_in_progress_to_pending()` before moving items
+  to the deferred list, so a task that was mid-run when the drain
+  fired still ends up on `/followup` instead of being orphaned.
+  `/stop` is separate: it moves `Pending|Blocked|InProgress` items
+  to deferred directly via its own `matches!` pattern
+  (kres-repl/src/session.rs), so a resumed REPL picks them up via
+  `/continue`.
+
 ### Skills
 - Loaded from `~/.kres/skills/*.md` at startup
 - Skill files scanned for absolute paths in backticks — referenced files pre-loaded
@@ -73,8 +101,8 @@ this repo's `configs/` tree:
 | `todo-agent.json` | Todo-list-maintenance agent (tools-disabled variant) |
 | `mcp.json` | MCP server definitions (installed only when semcode-mcp is available) |
 | `settings.json` | Per-user defaults (today: per-role model ids). CLI flags `--fast-model`, `--slow-model`, `--main-model`, `--todo-model` override the matching role; a known `--slow <tag>` (sonnet/opus) also overrides the slow model id unless `--slow-model` is given |
-| `prompts/*.system.md` | System prompts referenced by each agent config |
-| `prompts/bug-summary.md` | Bug-report template for `/summary` and `kres --summary` |
+| `system-prompts/*.system.md` | Optional operator overrides for agent system prompts. Default prompts are embedded in the kres binary (`kres-agents/src/embedded_prompts.rs`); a file at `~/.kres/system-prompts/<basename>` shadows the embedded copy. Empty by default |
+| `commands/<name>.md` | Optional operator overrides (or additions) for slash-command templates. Shipped commands `review`, `summary`, `summary-markdown` are embedded in the kres binary (`kres-agents/src/user_commands.rs`). A file at `~/.kres/commands/<name>.md` shadows the embedded copy; adding a new `<name>.md` creates a `/name` command invocable via `--prompt "name: extra"` or `--prompt "/name extra"`. Empty by default |
 | `skills/*.md` | Domain knowledge files |
 
 Rate limiters are shared across agents that use the same API key string.
@@ -85,9 +113,13 @@ Rate limiters are shared across agents that use the same API key string.
 |---------|--------|
 | `/tasks` `/task` | Show active tasks and states |
 | `/todo` | Show pending items (ready/blocked) + completed count |
+| `/plan` | Show the current plan + per-step status (produced by `define_plan`) |
+| `/resume [PATH]` | Load a persisted `session.json` (defaults to `<results>/session.json.prev` → live file). Overwrites in-memory state |
 | `/todo --clear` | Clear all todo items |
 | `/cost` | Token usage by agent role and model |
-| `/summary` | Main agent synthesizes accumulated findings |
+| `/summary [FILE]` | Fast agent renders the run's report.md + findings.json into a summary via the embedded `summary` slash-command template. Output defaults to `summary.txt` in the results dir. Auto-chunks findings when the prompt exceeds the fast agent's `max_input_tokens` and runs a combine pass to merge the partials |
+| `/summary-markdown [FILE]` | Same as `/summary` but uses the `summary-markdown` template and defaults the filename to `summary.md` |
+| `/review <target>` | Compose the embedded `review` slash-command template with `<target>` and submit as a new task — CLI equivalent of `--prompt 'review: <target>'` |
 | `/report <file>` | Write all findings to markdown file |
 | `/followup` | Show deferred items (identified but skipped when goal met) |
 | `/next` | Run next todo item |
@@ -148,9 +180,10 @@ Rate limiters are shared across agents that use the same API key string.
   prompts/                    # System prompts + bug-summary.md
   skills/                     # Skill files (kernel.md, …)
   sessions/<ts>/              # Per-run artifacts when --results not set
-    findings.json             # Cumulative findings (history in findings-N.json)
+    findings.json             # jsondb-backed canonical findings (delta-applied, no history)
     report.md                 # Append-only narrative
-    bug-report.txt            # Output of /summary or kres --summary
+    session.json              # Plan + todo + deferred + counters (resume state)
+    summary.txt               # Output of /summary or kres --summary (summary.md with --summary-markdown)
 
 .kres/logs/<session-uuid>/    # Next to cwd, one dir per REPL run
   code.jsonl                  # All fast + slow agent turns