@buildaharness/aielia 0.3.4 → 0.3.5

package/README.md

@@ -142,7 +142,7 @@ contradiction detection considers it:
 
 | Tier | Answers | Allowed `FactSource` | Retention | Contradiction-checked | Where it lives |
 |---|---|---|---|---|---|
-| **Episodic** | "What was said or mused, unconfirmed?" | any | session | no | `facts:<sessionId>` — also every `model_inferred`/`observed` `UserFact` lands here regardless of its `durable` bit; an unconfirmed model guess is never Knowledge |
+| **Episodic** | "What was said or mused, unconfirmed?" | any | session | no | `facts:<sessionId>`, plus `episodic:<sessionId>` session digests when `AUDIT_EPISODIC_DIGEST` is on (see below) — also every `model_inferred`/`observed` `UserFact` lands here regardless of its `durable` bit; an unconfirmed model guess is never Knowledge |
 | **Semantic** | "What's currently stated as true?" | `user_asserted`, `externally_verified` | durable (by tier policy) | yes | `facts:<sessionId>` or `facts:durable`, depending on the existing promotion policy (`durable` bit) — the tier a fact belongs to and whether it's actually promoted are tracked separately today |
 | **Identity** | "Who is this?" (name, "call me X") | `user_asserted` | durable | yes | `facts:durable` |
 | **Preference** | "What does the user want?" (stated preference) | `user_asserted` | durable | yes | `facts:durable`; a *configured* preference (backend, model, `enableShell`) is separately `AssistantConfig` (`config.ts`), not a `UserFact` at all |
@@ -158,6 +158,40 @@ the other half is that `isKnowledgeTier()` only returns true for `semantic`,
 `identity`, and `preference` — episodic entries, procedural weights, and
 commitments never enter contradiction detection as if they were beliefs.
 
+**Episodic memory now has a store and a writer, behind a flag.** With
+`AUDIT_EPISODIC_DIGEST` on (off by default; the cross-session scenario improved on 2 of 2 seeds, one scenario only), the assistant
+writes one short handoff digest per session (objective, what was done,
+decisions, open items, next step) at the session edge (`/new`, exit), never
+inside a turn, to `episodic:<sessionId>`. A second use of the same writer runs
+just before transcript compaction drops old messages, so facts in those
+messages are queued for confirmation instead of being lost. Digests are kept
+apart from durable facts and are never injected into the prompt: they reach the
+model only through the read-only `recall_memory` tool (`AUDIT_RECALL_TOOL`, off
+by default), which returns an index of recent digests and then one digest by id,
+wrapped as untrusted context. The model picks which digest to open; there is no
+keyword search. Digest text passes the same secret judgement as any other
+memory write, and a digest can never populate a Knowledge tier (the extended
+INV-16: a non-user origin is episodic only). Digests are kept for a retention
+window (default 90 days, `AUDIT_EPISODIC_RETENTION_DAYS`) and are covered by
+`/memory forget` and `/memory export`. This changes what `/new` keeps: durable
+facts as before, plus digests when the flag is on. Evidence is one scenario and two seeds; it also adds one model call per session edge.
+
+Related write-path controls: `AUDIT_MEMORY_BUDGETED_RENDER` is **on by default**
+(character-budgeted, priority-ordered facts block with keyed supersession, so a
+corrected fact replaces the old one; `=0` restores the old 20-fact slice; budget
+`memoryBudgetChars`, default 4000). In a six-week corrections scenario it fixed a
+stale-fact failure on 3 of 3 seeds. Off by default: `AUDIT_MEMORY_WRITE_GATE`
+(semantic secret redaction and instruction-shaped-text flagging before a fact is
+promoted; a missing judgement means the fact is not promoted; no outcome
+difference was seen because the base model already resisted the planted secret),
+`AUDIT_MEMORY_AUDIT_LOG` (`/memory history`, `/memory undo <seq>`),
+`AUDIT_MEMORY_REVIEWER` (post-turn reviewer; its writes are staged for
+`/memory confirm` unless `memoryWriteMode` is `auto`) and
+`AUDIT_MEMORY_CONSOLIDATION` (staged merge and archive proposals; not measured as
+an outcome). Known gaps: the model sometimes keys the same attribute two ways, so
+a correction can miss the old fact, and a natural-language "forget that" request
+is staged for approval rather than applied.
+
 The conversation transcript itself (`transcript:<sessionId>`, "what was
 said" verbatim) and a single turn's `AnswerClaim` (evidence vs. claim for one
 reply, see below) sit outside this tier system — they're the raw material a
@@ -448,6 +482,9 @@ without ever exposing the filesystem or shell.
 
 ## Batch research budget
 
+> The list detector below is a lexical check (the `batch-list` family), **off by default**: set
+> `ASSISTANT_LEXICAL_ON=batch-list` (or `ASSISTANT_LEXICAL_MODE=enabled`) to use batch research.
+
 The flat `maxSteps` cap that governs an ordinary chat turn (see below) doesn't
 know the difference between a one-question turn and a turn asking for the same
 lookup across many items — a 7-item batch and a 1-item question get the same
@@ -705,17 +742,7 @@ non-functional `web_search`. The active state is shown in the startup banner
 running session with `/config set` instead of an env var — see
 "Configuration" below.
 
-Set `ASSISTANT_LLM_BACKEND=claude-cli` to skip the proxy entirely and run turns
-through a local `claude -p` subprocess instead, using your already-authenticated
-Claude Code CLI session rather than an API key (`CLAUDE_PATH` overrides the
-`claude` binary path if it's not on `PATH`):
-
-```bash
-ASSISTANT_LLM_BACKEND=claude-cli npm run cli --workspace=packages/aielia
-```
-
-Or skip both the proxy and claude-cli and call a provider directly with your
-own API key — `ASSISTANT_LLM_BACKEND=anthropic|openai|openrouter` plus
+Skip the proxy and call a provider directly with your own API key — `ASSISTANT_LLM_BACKEND=anthropic|openai|openrouter` plus
 `ASSISTANT_API_KEY`:
 
 ```bash
@@ -803,6 +830,27 @@ package's persistence, it's a real file, not encrypted, so `authToken` and
 `braveApiKey` are stored in plaintext there. This is the same trust boundary
 the repo's root `.env` already has, not a new one.
 
+### Lexical checks (off by default)
+
+Every lexical pass in this package (regex / keyword / phrase-list checks over natural-language text) sits
+behind one switch, `lexicalMode`, and **every one is off by default** in the CLI, the browser tab and the
+desktop app — the LLM-based checks do the work instead. The ten families (`fact-markers`, `coding-fact`,
+`injection`, `enumeration`, `risk`, `task-cancel`, `plan-mode`, `batch-list`, `tool-yield`,
+`template-keywords`) and what "off" means for each are listed in `src/lexical/lexical-mode.ts`.
+
+- `ASSISTANT_LEXICAL_MODE=enabled` turns every family back on (the pre-2026-09-28 behaviour).
+- `ASSISTANT_LEXICAL_ON=batch-list,task-cancel` turns only those families back on (`all` for every one).
+- `ASSISTANT_LEXICAL_OFF=<families>` turns families off even under `enabled`; it wins over ON.
+
+The mode is handed to `@buildaharness/harness`'s own lexical switch at the start of every run, so
+`enabled` turns the harness's checks back on too (see that package's README).
+
+Some families have no LLM replacement, so with them off the feature they drive does not run: batch
+research (`batch-list`), matching "skip that step" to one step of an active plan (`task-cancel`), leaving
+plan drafting by phrase (`plan-mode`; `/plan` still works), and the batch sub-loop's lexical dead-end
+detection (`tool-yield`; the web tool's own "No results found." still counts). With `risk` off, a durable
+plan step saved without its own risk level is treated as HIGH.
+
 ## REPL commands
 
 Type `/help` inside a running CLI session for this list. All of them read or
@@ -820,10 +868,11 @@ change local session/config state — none of them make an LLM call themselves.
 | `/search <query>` | Ranked search over past messages, across every session in this install — a hit is the one message that matched, not the whole session transcript around it. Read-only: never an LLM call or network request. Scoring is tokenized/graduated, not exact-substring-only (see "Conversation history" above), and a query matching nothing returns an explicit "no results" line |
 | `/model [name]` | Show the active model, or switch it — a thin alias over `/config set model <name>` (see "Configuration" above); rejected the same way if `model` is pinned by `ASSISTANT_MODEL` |
 | `/cost` | Show token usage for the last turn and the running session total |
-| `/doctor` | Check proxy reachability (proxy backend) or the `claude` binary (claude-cli backend), plus workspace root and data dir health — no dedicated check yet for the anthropic/openai/openrouter backends, only the backend-agnostic checks run for those |
+| `/doctor` | Check proxy reachability (proxy backend), plus workspace root and data dir health — no dedicated check yet for the anthropic/openai/openrouter backends, only the backend-agnostic checks run for those |
 | `/why` | Explain the harness path the last turn took (verification confidence + node sequence) |
 | `/sources` | List files/URLs the last turn actually consulted |
 | `/plan` | Show the active structured plan's task status |
+| `/plan graph [thread-id]` | Draw the active plan as a read-only dependency graph (needs `planGraphMode` enabled; full-screen pane in the terminal UI with arrows/Tab/Enter/q, one static render in the plain REPL). Glyphs, mapping and keys: `docs/plan-visualization.md` |
 | `/checkpoint [clear]` | Inspect a stuck in-progress harness checkpoint (step, node, failed-resume count so far), or `/checkpoint clear` to discard it — see "Recovering a stuck checkpoint" above. Scoped to just the checkpoint: unlike `/clear`, transcript/facts/plan are untouched |
 
 ### `/cost` and real vs. estimated dollar figures
@@ -831,13 +880,7 @@ change local session/config state — none of them make an LLM call themselves.
 Token counts are always real, on every backend. The dollar figure attached to
 them is not always the same kind of number:
 
-- **claude-cli backend**: `costUsd` comes straight from `claude
-  --output-format json`'s own `total_cost_usd` field — real Anthropic
-  accounting. It may read `$0` if the underlying `claude` session is
-  authenticated against a Pro/Max subscription rather than API billing, in
-  which case `$0` does *not* mean "this turn was free" — `/cost`'s output
-  says so explicitly.
-- **every other backend** (`proxy`, `anthropic`, `openai`, `openrouter`):
+- **every backend** (`proxy`, `anthropic`, `openai`, `openrouter`):
   none of these compute a dollar cost themselves — only the raw token counts
   each provider's response already includes (`@buildaharness/proxy` is a thin
   pass-through, see `packages/proxy/src/forward.ts`; the three direct clients

package/dist/PlanGraphPane.d.ts

@@ -0,0 +1,19 @@
+import { VizNode } from './plan-viz/types.js';
+/** Width at which the detail drawer docks to the right instead of being toggled. */
+export declare const DRAWER_DOCK_COLUMNS = 120;
+export interface PlanGraphPaneProps {
+    nodes: readonly VizNode[];
+    columns: number;
+    rows: number;
+    /** False while an approval prompt owns the keyboard: the pane stays drawn but ignores keys. */
+    active?: boolean;
+    color?: boolean;
+    ascii?: boolean;
+    onClose: () => void;
+}
+/**
+ * Read-only plan graph pane (plans/plan_visualization_plan.html, R2). Keys: arrows (or [ ]) move along edges and ranks, j/k and PgUp/PgDn scroll the detail drawer,
+ * Tab/Shift+Tab walk task order, Enter or d toggles the detail drawer below 120 columns, q or Esc closes.
+ * No key edits anything: the plan changes only through the approval gate.
+ */
+export declare function PlanGraphPane({ nodes, columns, rows, active, color, ascii, onClose }: PlanGraphPaneProps): React.JSX.Element;