HodeTech · cemililik · Aug 25, 2026 · Aug 24, 2026 · Aug 25, 2026
diff --git a/AGENTS.md b/AGENTS.md
@@ -24,7 +24,9 @@ gateway; engine stays local, Phase 2) — split across build phase 5 (managed in
 phase 6 (cloud execution + portal); the engine is identical across all three (ADR-0012..0015).
 **Status: Phase 1 is complete; Phase 2 (CLI) is feature-complete (v0.1.1 release cut, publish
 pending); Phase 2.5 (CLI Consolidation) is complete (M2.5-4, 2026-07-08); Phase 2.6
-(Conversational Authoring and the First-Class CLI) is next up.**
+(Conversational Authoring and the First-Class CLI) is in progress, with an interlude —
+Phase 2.6.5 (Core reliability remediation) — running between remediation Waves 1 and 2:
+14 of 47 items closed, including all eight P0 blockers (ADR-0078–ADR-0084, merged 2026-08-24).**
 See [docs/roadmap/current.md](docs/roadmap/current.md) for live status.
 
 ## The non-negotiable rules

diff --git a/CLAUDE.md b/CLAUDE.md
@@ -55,6 +55,13 @@ PR #75, 2026-07-13). The phase is a full-screen Home-managed CLI with conversati
 browsers, competitor-breadth tools, settings/theming/`en`+`tr` localization, and the
 run-ops resume follow-up.
 
+**An interlude is running between Wave 1 and Wave 2 of the remediation: Phase 2.6.5 (Core reliability),
+14 of 47 items closed** — `W0` (PR #82, 2026-08-11) and `W1`, the eight P0 blockers plus `CR-92`, merged
+2026-08-24 (PR #83) behind [ADR-0078](docs/decisions/0078-ordered-durable-append-and-the-terminal-outbox.md)–[ADR-0084](docs/decisions/0084-consent-before-a-local-mcp-spawn.md):
+ordered durable append, cross-process run ownership, the durable effect journal, untrusted compaction summaries,
+the stream-grammar seam obligation, engine-side input admission and resume identity, and consent before a local
+MCP spawn. No new product surface — only the invariants an existing surface already claims.
+
 For live status, per-PR history, milestone dates, and open obligations, see the canonical
 home [docs/roadmap/current.md](docs/roadmap/current.md);
 [README.md](README.md) is the public overview;

diff --git a/docs/decisions/0069-string-width-for-the-cli-renderer.md b/docs/decisions/0069-string-width-for-the-cli-renderer.md
@@ -2,7 +2,7 @@
 
 - **Status**: Accepted
 - **Date**: 2026-07-10 (accepted 2026-07-11)
-- **Related**: [ADR-0068](0068-full-screen-tui-renderer-ink7-harness.md) · [ADR-0047](0047-cli-render-seam-and-framework-free-cores.md) · [ADR-0067](0067-node-supported-floor-22-reaffirm-better-sqlite3.md)
+- **Related**: [ADR-0068](0068-full-screen-tui-renderer-ink7-harness.md) · [ADR-0047](0047-cli-framework-commander-ink-clack.md) · [ADR-0067](0067-node-supported-floor-22-reaffirm-better-sqlite3.md)
 
 > **Accepted on 2026-07-11** when the maintainer merged PR #74 (2.6.F) into `main` — the merge carries the
 > `string-width` dependency, which is the [CLAUDE.md](../../CLAUDE.md) rule 2 approval this ADR gated. The code in

diff --git a/docs/reference/shared-core/database-schema.md b/docs/reference/shared-core/database-schema.md
@@ -210,7 +210,7 @@ erDiagram
 
 #### `llm_providers`
 
-Registered LLM providers. The actual API key never lives here — only a reference; the key is stored in the OS keychain (see [keychain-and-secrets.md](keychain-and-secrets.md)).
+Registered LLM providers. The actual API key never lives here — only a reference; the key is stored in the OS keychain (see [keychain-and-secrets.md](../desktop/keychain-and-secrets.md)).
 
 | Column | Type | Constraints |
 |--------|------|-------------|
@@ -849,7 +849,7 @@ This realizes the concurrent-process write requirement recorded in the [ADR-0064
 
 At-rest encryption of `history.db` is **per-surface**:
 
-- **Desktop:** opened with SQLCipher. The passphrase is derived from a stable machine secret (combined with the OS keychain entry) so the database opens on restart without prompting the user; see [keychain-and-secrets.md](keychain-and-secrets.md).
+- **Desktop:** opened with SQLCipher. The passphrase is derived from a stable machine secret (combined with the OS keychain entry) so the database opens on restart without prompting the user; see [keychain-and-secrets.md](../desktop/keychain-and-secrets.md).
 - **CLI (Phase 2):** opened with `better-sqlite3` **unencrypted**, guarded by owner-only OS file permissions — `~/.relavium/` at `0700` and `history.db` (with its `-wal`/`-shm` sidecars) at `0600`, set with an explicit `chmod` (umask-independent, applied even to a pre-existing directory). On Windows, POSIX mode bits do not apply (`chmod` is a no-op); protection falls to the per-user `%USERPROFILE%` NTFS ACL. The file holds **no credentials** — keys stay in the OS keychain ([ADR-0006](../../decisions/0006-os-keychain-for-api-keys.md)) and the engine masks secrets at the bus before persistence ([ADR-0036](../../decisions/0036-run-loop-substrate-event-bus-and-execution-host.md)) — so the unencrypted-at-rest content is run data (prompts, outputs, costs), not secrets. Rationale and the cross-surface Phase-3 follow-on: [ADR-0050](../../decisions/0050-cli-history-db-at-rest-posture.md).
 
 The per-project `runs.db` is **not** encrypted on any surface because it is intentionally git-committed and contains only non-sensitive run metadata (no prompts, completions, or tokens).

diff --git a/docs/roadmap/current.md b/docs/roadmap/current.md
@@ -90,7 +90,7 @@ flowchart TD
  W0["Wave 0 — One true baseline<br/>baseline ✅ · CI truth · numbers"]
  W1["Wave 1 — Stop the bleeding ✅<br/>3 CRITICALs · cost cap · ADR-0074"]
  LEDGER["#W15-1 — realized-cost ledger ✅<br/>ADR-0076 + ADR-0077"]
- P265["Phase 2.6.5 — Core reliability<br/>46 CR items · 8 P0 ADRs · 9 closed<br/>absorbs the hostile-MCP class"]
+ P265["Phase 2.6.5 — Core reliability<br/>47 CR items · 8 P0 blockers behind ADR-0078–0084<br/>W0+W1 closed (14)<br/>absorbs the hostile-MCP class"]
  W2["Wave 2 — Shut the doors<br/>fs jail · secrets · config trust<br/>certifies 2.5.5 EXIT 1–3"]
  W3["Wave 3 — Clear the ground<br/>god-file decomposition · CLI net"]
  W4a["Wave 4a — The spine<br/>2.6.A/D/H/K + 2 ADRs"]
@@ -464,16 +464,23 @@ gaps**: effect journal, stdio MCP consent-before-spawn, run lease, compaction tr
 input admission, event-log ordering. Three separate reviews landing on the same seven points is not opinion.
 
 The full, self-contained work list is
-[phase-2.6.5-core-reliability-remediation.md](phases/phase-2.6.5-core-reliability-remediation.md) — **46 items**
+[phase-2.6.5-core-reliability-remediation.md](phases/phase-2.6.5-core-reliability-remediation.md) — **47 items**
 (`CR-01`…`CR-95`) with evidence, fix, acceptance criteria and a decision/ADR/gate register, written so the work
 can be done from that document alone. An adversarial plan review on 2026-08-10 corrected the phase boundary,
 the exit rule and the execution order, and added two items (`CR-17` resume identity, `CR-63` `input_schema`
-docs-only).
+docs-only); `CR-64` came from the Batch 1 triage and `CR-21b` from ADR-0082 §10, which is why the total has
+moved since the list was first written.
 
-> **Live status — 6 of 46 closed, merged 2026-08-11 (PR #82).** The prerequisite (`#W15-1`), the oracle
-> (`CR-90`, `CR-91`) and all of `W0` (`CR-01`–`CR-03`). `CR-64` was added in the same batch and is open.
-> **Next is the durability spine — `CR-10` first**, and it is ADR-first: no code until the decision is
-> recorded. Per-item history and the carried-forward gaps live in the phase document.
+> **Live status — 14 of 47 closed. `W0` and `W1` are both done.**
+>
+> - **Batch 1, merged 2026-08-11 (PR #82)** — the prerequisite (`#W15-1`), the oracle (`CR-90`, `CR-91`) and
+> all of `W0` (`CR-01`–`CR-03`). `CR-64` was added in the same batch and is open.
+> - **Batch 2, merged 2026-08-24 (PR #83)** — all of `W1`: the eight P0 blockers plus `CR-92`, behind
+> ADR-0078…ADR-0084. A comprehensive review of the assembled PR found seven further defects, six of them in
+> the W1 code itself; all seven were fixed and mutation-verified before merge.
+>
+> **Next is `W2` — liveness and deadlines.** Per-item history, the seven post-review findings and the
+> carried-forward gaps live in the phase document.
 
 **This is the corrected execution order, and it is what the graph above shows:**
 

diff --git a/docs/roadmap/deferred-tasks.md b/docs/roadmap/deferred-tasks.md
@@ -319,6 +319,54 @@ Severity is the review's verified rating. Check an item off in the PR that resol
  there is concrete surface demand or telemetry showing operators need an earlier signal.
  *(1.AC; ADR-0028; config-spec.md; workflow-yaml-spec.md)*
 
+## Phase 2.6.5 `W1` residuals (PR #83, merged 2026-08-24)
+
+Named rather than left implicit, because every one of them was examined during `W1` and consciously left
+open — an unstated residual reads as an oversight, and the review that found the seven `PR83-*` defects said
+so directly: these "should remain visible rather than disappear behind the green test suite."
+
+- [ ] **Terminal-outbox drain, callable from any surface ([ADR-0078](../decisions/0078-ordered-durable-append-and-the-terminal-outbox.md) §4/§5).**
+ `drainTerminalOutbox` is a `WorkflowEngine` method, and `run`/`gate` are the only commands that construct
+ one — `chat`, `agent run` and the bare-invocation Home run on `AgentSession` and have no engine to drain
+ with. `PR83-06` corrected the documentation and made `relavium status` NAME a run whose terminal is held
+ (`terminalHeld` in `--json`), but the drain itself is still reachable from two commands only. The real fix
+ lifts it off the engine onto its host ports (`store` + `terminalOutbox` + `runLeases` + `mediaReferences`),
+ so any surface can call it at startup. **Acceptance:** a held terminal is retried before `status` reads the
+ run; a drain failure stays non-fatal and preserves the entry; the entry is removed only once the exact
+ terminal is durably present. Deliberately NOT done inside `status` as a read command — draining claims a run
+ lease, and a status read must not take ownership of a run another process may be finishing.
+ *(packages/core/src/engine/engine.ts; apps/cli/src/commands/status.ts; ADR-0078 §4/§5)*
+
+- [ ] **Check-then-write symlink window on the append-only files.** The MCP grant store and the terminal
+ outbox both `lstat` a path and then append to it by name, and their comments admit the window. On POSIX a
+ descriptor-based open with `O_NOFOLLOW | O_APPEND | O_CREAT` followed by `fstat`/`fchmod` closes it. Recorded
+ as defense-in-depth rather than a blocker: the present threat already requires write access to the user's
+ `~/.relavium` directory. *(apps/cli/src/engine/{mcp-consent,terminal-outbox}.ts)*
+
+- [ ] **Consent overflow can be approved without being read ([ADR-0084](../decisions/0084-consent-before-a-local-mcp-spawn.md) §7).**
+ The prompt shows 12 arguments / environment entries and states how many more exist, asking the user to open
+ the artifact. That bounds terminal abuse well, but a declaration whose dangerous part sits in the hidden tail
+ can still be approved. A premium trust experience would require an explicit view/acknowledge step when
+ overflow exists. *(apps/cli/src/mcp/consent-prompt.ts; ADR-0084 §7)*
+
+- [ ] **The effect audit's occurrence is coarser than its type promises ([ADR-0080](../decisions/0080-durable-effect-journal-and-the-tiered-effect-contract.md)).**
+ `effect-journal-store.ts` records openly that the provider-attempt and provider tool-call identity are not
+ threaded to dispatch. Nothing load-bearing depends on them — the dedup key is the identity and the resume
+ gate reads the scope — but an operator-facing journal should eventually carry the exact occurrence
+ `EffectAttemptId` describes. *(packages/db/src/effect-journal-store.ts)*
+
+- [ ] **An authored `pattern` with no `max_length` can still stall the process ([ADR-0083](../decisions/0083-input-admission-and-a-resume-that-verifies-its-own-identity.md) §4).**
+ `max_length` bounds the input a catastrophic regex chews on, and it is checked first — but only when the
+ author supplied it. ADR-0083 accepts this floor explicitly. A stronger contract would require `max_length`
+ whenever `pattern` is present, or admit only a demonstrably safe regex subset. *(packages/shared/src/workflow.ts; ADR-0083 §4)*
+
+- [ ] **Two SonarCloud CRITICALs are declines, not debt — mark them Accepted in the UI.**
+ `WorkflowEngine#emitDurable`'s closure (23) and `dispatch` (22) are both over the cognitive-complexity
+ threshold and neither can be split without harm: in both, the branches ARE the ordering the surrounding
+ comments exist to protect. `#emitDurable`'s extraction would reinsert the microtask hop that reordered the
+ log once already. Resolving them requires a maintainer action in SonarCloud, not a code change.
+ *(packages/core/src/engine/engine.ts; packages/core/src/tools/registry.ts)*
+
 ## Interpolation engine (1.L2) follow-ups
 
 > A comprehensive multi-dimensional pre-merge review of **1.L2** (PR #15, merged 2026-06-12) confirmed

diff --git a/docs/roadmap/phases/phase-2.6.5-core-reliability-remediation.md b/docs/roadmap/phases/phase-2.6.5-core-reliability-remediation.md
@@ -1,7 +1,8 @@
 # Phase 2.6.5 — Core reliability remediation (interlude)
 
-- **Status**: in progress — the prerequisite and the oracle are closed; the durability spine is next
-- **Opened**: 2026-08-09 · **Plan corrected**: 2026-08-10 · **First batch merged**: 2026-08-11 (PR #82)
+- **Status**: in progress — **`W0` and `W1` are closed** (14 of 47 items); `W2` is next
+- **Opened**: 2026-08-09 · **Plan corrected**: 2026-08-10 · **First batch merged**: 2026-08-11 (PR #82) ·
+ **`W1` merged**: 2026-08-24 (PR #83)
 - **Predecessor**: Wave 1 of the 2.5.5 remediation (complete — PR #81), then the `#W15-1` realized-cost
  ledger implementation (**complete 2026-08-10**, ADR-0076 + ADR-0077 — see [Prerequisite](#prerequisite))
 - **Successor**: Wave 2 of the 2.5.5 remediation, **reduced** — this phase absorbs Wave 2's hostile-MCP
@@ -120,8 +121,46 @@ shape.
 >
 > `CR-64` was **added** in the same batch (from the YAML/git-native review triage); it is open.
 >
-> **Next: the durability spine**, `CR-10` → `CR-11` → `CR-92` → `CR-12`, in that order and ADR-first. The
-> oracle exists specifically to prove it, and `CR-10` is the item everything else assumes.
+> **Next after Batch 1 was the durability spine**, `CR-10` → `CR-11` → `CR-92` → `CR-12` — a dependency
+> chain, recorded here as the history of why that order was chosen. Batch 2 below is the whole of `W1`: those
+> four plus the independent `CR-13`, `CR-14`, `CR-15`, `CR-16` and `CR-17` lines, and its table is ordered by
+> ADR rather than by the chain.
+>
+> **Batch 2 — `W1`, merged to `main` 2026-08-24 (PR #83).** All eight P0 blockers plus `CR-92` — nine items
+> behind **seven** ADRs, since `CR-92` shares `CR-10`'s and `CR-15`/`CR-17` share one. Each item had an Opus
+> and a Sonnet review round folded before the next one started.
+>
+> | Item | Closed | ADR |
+> |------|--------|-----|
+> | `CR-10` | 2026-08-11 | [ADR-0078](../../decisions/0078-ordered-durable-append-and-the-terminal-outbox.md) |
+> | `CR-92` | 2026-08-11 | with `CR-10`'s |
+> | `CR-11` | 2026-08-17 | [ADR-0079](../../decisions/0079-cross-process-run-ownership-lease-and-fencing-token.md) |
+> | `CR-12` | 2026-08-18 | [ADR-0080](../../decisions/0080-durable-effect-journal-and-the-tiered-effect-contract.md) |
+> | `CR-13` | 2026-08-18 | [ADR-0081](../../decisions/0081-the-compaction-summary-is-untrusted-and-the-system-prompt-is-branded.md) |
+> | `CR-14` | 2026-08-19 | [ADR-0082](../../decisions/0082-the-stream-grammar-is-a-seam-obligation-and-every-attempt-has-a-deadline.md) |
+> | `CR-15` · `CR-17` | 2026-08-19 | [ADR-0083](../../decisions/0083-input-admission-and-a-resume-that-verifies-its-own-identity.md) |
+> | `CR-16` | 2026-08-20 | [ADR-0084](../../decisions/0084-consent-before-a-local-mcp-spawn.md) |
+>
+> **A comprehensive review of the assembled PR found seven further defects, all fixed before merge** — and
+> they are recorded here rather than quietly folded in, because six of the seven were defects in the W1 code
+> *itself*: this batch's own mechanisms failing the guarantees they were written to establish. Three were
+> reproduced with executable counterexamples before any fix was written.
+>
+> | Id | Severity | What was wrong |
+> |----|----------|----------------|
+> | `PR83-01` | High | The compare-and-append guard checked only that the log's max EQUALLED the caller's belief, never that the incoming event was AHEAD of it. Sequence gaps are legitimate, so a stale terminal's number is both unique and lower: one appended behind durable work, `applyDerived` marked the run finished, and the outbox drain deleted its recovery entry on that false success. |
+> | `PR83-02` | High | `JSON.stringify` deletes a property whose value is `undefined`, so a legitimately-`undefined` tool result replayed as the envelope's own metadata object. The in-memory journal held results BY REFERENCE, which is why no core test saw it. |
+> | `PR83-03` | High | A caller cancellation landing before `openDeadline` was forwarded to the provider controller but never latched, so `race()` had nothing to observe and waited out the 120 s deadline against a provider that ignores its signal. |
+> | `PR83-04` | Medium | A proven non-dispatch (missing host capability) left its row `prepared` forever — unresolved, resume-blocking, never swept. The existing test asserted that state directly under a comment saying the point was to avoid it. |
+> | `PR83-05` | Medium | SQLite `settle` discarded the `changes` count, so a missing or already-terminal row reported durable success for an effect that may have landed. |
+> | `PR83-06` | Medium | Exit 5 told users recovery happens "on the next `relavium` start"; only `run` and `gate` drain the outbox, so `status` could never resolve it. |
+> | `PR83-07` | Low | `database-schema.md` said effect retention was unimplemented while this PR ships both sweeps, and drew the run→lease edge as exactly-one for a row that is created on acquire and deleted on release. |
+>
+> The **six code defects** are mutation-verified — each test was confirmed to FAIL with its fix reverted.
+> `PR83-07` is a documentation correction with nothing executable to mutate; it was verified against the
+> shipped `effect-retention.ts` and the lease row's actual lifecycle. Two additional
+> coverage gaps surfaced that way and are now pinned: the fold-failure path's `contentCommitted` stamp, and
+> the guard that omits the `mapped` projection when a node configured no `output_mapping`.
 >
 > **Carried forward, named rather than implied.** ADR-0077's required regression (a ledger write refused while
 > a sibling's `#failure` already suppressed the abort) is unbuilt, and `#runAttempt`'s money-durability arm is
@@ -841,6 +880,13 @@ mark.** Wave 1's completion claim was wrong twice before this discipline was ado
 caught it a third time: `CR-14` and `CR-92` were both implemented and both still carried an OPEN heading here
 (*"needs an ADR"*, *"in the durability spine"*), which is precisely the failure the criterion exists to catch.
 
+**And a fourth time, from the outside.** A comprehensive review of the assembled PR found six defects in the
+code this register vouches for — the mechanisms of `CR-10`, `CR-12` and `CR-14` failing the guarantees they
+were written to establish (see `PR83-01`…`PR83-06` in [Progress](#progress)). Every one is fixed and
+mutation-verified, and the register rows below name the tests that now hold them. The lesson is the one the
+criterion already encodes, sharpened: reading the code you wrote is necessary and not sufficient, because the
+reader shares the author's assumptions. Three of the six were only settled by an executable counterexample.
+
 | Item | ADR | The code that closes it | The test that would fail if it were reverted |
 |------|-----|--------------------------|-----------------------------------------------|
 | `CR-10` | — | `packages/core/src/engine/append-audit.ts` — a `RunStore` decorator recording asks, commits and outcomes per run | `append-audit.test.ts` + the certification against a real `history.db`. The property is **not** a log assertion: streamed events take sequence numbers and are never persisted, so a healthy run's log reads `[0,1,2,3,5,10,…]` and a lost event is byte-identical to a skipped one. The audit supplies the witness the log cannot. |