feat: prepare beta16 memory and provider evolution

.github/workflows/codeql.yml

@@ -48,6 +48,13 @@ jobs:
     name: Analyze JavaScript and TypeScript
     runs-on: blacksmith-4vcpu-ubuntu-2404
     timeout-minutes: 20
+    # Reuse main's CodeQL database for pull requests and analyze the diff with
+    # the existing code as context. CodeQL falls back to a full scan when the
+    # base database, diff ranges, or runner requirements are unavailable.
+    env:
+      CODEQL_ACTION_OVERLAY_ANALYSIS: "true"
+      CODEQL_ACTION_OVERLAY_ANALYSIS_CODE_SCANNING_JAVASCRIPT: "true"
+      CODEQL_ACTION_DIFF_INFORMED_QUERIES: "true"
     steps:
       - name: Checkout repository
         uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10

CHANGELOG.md

@@ -1,5 +1,32 @@
 # Changelog
 
+## 0.1.44-beta.16 - 2026-09-23
+
+### Added
+
+- Add opt-in Memory & Evolution setup controls with bounded evidence, policy
+  gating, candidate review, and reversible project-scoped promotion.
+- Add adaptive reasoning policy/runtime support with explicit user controls and
+  focused validation coverage.
+
+### Changed
+
+- Accept provider-native and local model identifiers without a rigid catalog
+  allowlist while preserving provider and path safety validation.
+- Aggregate prompt-cache usage across the active user turn for more accurate TUI
+  status reporting.
+
+### Fixed
+
+- Harden queued peer delivery, compaction recovery, and session state checks so
+  accepted messages are not persisted into a target session while its runner is
+  still busy.
+
+### Tests
+
+- Add focused regressions for Memory & Evolution, adaptive reasoning, provider
+  compatibility, model validation, prompt cache reporting, and peer delivery.
+
 ## 0.1.44-beta.15 - 2026-09-14
 
 ### Fixed

docs/memory-center.md

@@ -33,6 +33,52 @@ Generated changes can come from normal extraction, Memory Center side chat, or
 Dream. They all land in the same review model: a proposal must be inspected,
 edited, applied, or rejected by the user before it changes saved memory.
 
+## Memory & Evolution (local implementation)
+
+Setup exposes project-scoped Evolution separately from **using saved memory**.
+Without explicit Evolution consent, existing legacy memory learning remains unchanged.
+Adopting Evolution replaces automatic legacy extraction/Dream for that project;
+it does not disable manually requested memory tools or delete saved entries.
+
+- **Off:** blocks new capture, inference and promotion; cancels Evolution work and
+  rejects late results. Previously approved artifacts remain active.
+- **Observe:** retains bounded metadata only, without free text or provider calls.
+- **Suggest:** authorized evidence produces review-only memory, skill and workflow
+  proposals. Skills/workflows require explicit approval; saving a workflow never runs it.
+- **Auto-safe (limited):** permits only an exact user correction in the form
+  `Project language: TypeScript.` (also JavaScript, Python, Rust or Go), with low
+  sensitivity and the project-stack category. Project memory must be empty, and
+  contradictory current corrections block automatic application. Everything else,
+  including permission/security/model rules, remains pending for review.
+
+Corrections can be entered from Setup or through a chat message beginning with
+`Correction:` or `Corrección:`. Optional tool/test evidence records host completion
+metadata, not stdout or assistant claims. A zero process exit status is not proof
+that an entire task passed a semantic audit.
+
+Provider processing requires separate consent. Local storage **does not mean local
+inference**: the selected provider may receive authorized evidence and consume quota.
+Evolution uses the configured extractor/distiller role without silently switching
+models. Evidence is limited to 500 records, 4 KiB per record and a 30-day eligible
+window; each run uses at most five current evidence records, lasts at most 60 seconds,
+and produces at most five candidates. Expired records are excluded from processing;
+physical pruning occurs when the evidence ledger is next rewritten.
+
+Daily execution uses the existing Dream service and registered workspaces. Setting a
+schedule does not install or start a service. There is a five-minute admission window,
+at most one claimed attempt per project/local date, and no automatic retry or late
+catch-up after failure or restart. The scheduler admits at most one Evolution run per
+tick. Use **Run once** for missed windows.
+
+**Revert Evolution memories** archives an unchanged newly-created memory and retires
+its graph projection without deleting graph content. An entry modified after promotion
+requires manual review instead. Reversion remains available while Evolution is Off.
+Capability rollback similarly checks the promoted file hash or workflow revision and
+refuses to overwrite later edits.
+
+This is contextual learning, not weight training or automatic source-code editing.
+Schema validation alone does not establish the usefulness of a generated capability.
+
 ## Memory Side Agent
 
 The side chat is intentionally constrained, but it is still powerful inside the

src/mendcode/packages/opencode/script/queue-compaction-smoke.ts

@@ -357,7 +357,7 @@ async function main() {
           "Press Enter to submit it, press Esc once to arm interruption, then press Esc again within five seconds to cancel it.",
           `The local model holds that request for ${Math.round(HOLD_MS / 1_000)} seconds so the queued state is visible.`,
           "Only one smoke can run at a time. Exit with /exit or Ctrl+C; child processes and the lock are cleaned automatically.",
-          "Expected result: the second Esc issues one cancellation; the compaction summary terminates once, Snake disappears, the panel collapses, and queued messages remain paired with their queued/send state.",
+          "Expected result: the second Esc issues one cancellation; the compaction summary terminates once, Snake's final board stays visible and stops advancing, the transcript collapses, and queued messages remain paired with their queued/send state.",
           "",
         ].join("\n"),
       )

src/mendcode/packages/opencode/src/cli/cmd/tui/component/prompt/index.tsx

@@ -111,6 +111,7 @@ import {
   readPromptStatusScript,
   resolvePromptCachePercent,
   resolvePromptStatus,
+  resolvePromptTurnCachePercent,
   type MendPromptStatusBuiltin,
   type MendPromptStatusScriptOutput,
   type MendPromptStatusScriptResult,
@@ -1749,6 +1750,11 @@ export function Prompt(props: PromptProps) {
     const sessionID = props.sessionID
     if (!sessionID) return
     const messages = sync.data.message[sessionID] ?? []
+    const turnPercent = resolvePromptTurnCachePercent({
+      messages,
+      activeAssistantID: findActiveWorkingAssistant()?.id,
+    })
+    if (turnPercent !== undefined) return turnPercent
     const active = findActiveWorkingAssistant()
     if (active?.liveUsage) return resolvePromptCachePercent(active.liveUsage)
     const last = messages.findLast((item): item is AssistantMessage => item.role === "assistant")