Harden local decision advice, fix Claude 5.x connections and document hook routing upgrade

.claude-plugin/skill-assets.sha256

@@ -3,4 +3,4 @@ c5d0c26f28c9ee14092f9deaf24c98dd8bef49d971fef2b7a537ffb1ab9f2887  .claude-plugin
 aeee7a94671ceb306fe2d24c5acc9f2d96ad8a8e7410536566799eea6265f080  skills/engraphis-memory/SKILL.md
 055655db84af07561d002f0c69744313d8413c39f3e873f941f0fa0b1e76dc66  skills/engraphis-memory/references/CONVENTIONS.md
 9d090a03f5b3f36a34d91f66b72c3844591f6915755ac3a6c6ba5f1b16977de5  skills/engraphis-memory/references/SCOPING.md
-210e42da31fc68f41c83b3ab9eb1f91a415dbc1ef148f75b6b170bddc35d5752  skills/engraphis-memory/references/TOOLS.md
+07de31349fc135895377cfcf852c490f732c19e83a627c311d724c5d2146dbb5  skills/engraphis-memory/references/TOOLS.md

.env.example

@@ -196,7 +196,7 @@ ENGRAPHIS_RETENTION_SUPERVISOR=none
 ENGRAPHIS_LLM_PROVIDER=openai
 # Model name (provider-specific):
 #   openai:      gpt-4o-mini, gpt-4o, gpt-4.1-mini, o4-mini ...
-#   anthropic:   claude-3-5-haiku-20241022, claude-3-5-sonnet-20241022 ...
+#   anthropic:   claude-sonnet-5-5, claude-opus-5-5, claude-haiku-4-5 ...
 #   google:      gemini-1.5-flash, gemini-2.0-flash ...
 #   openrouter:  anthropic/claude-3.5-sonnet, openai/gpt-4o-mini ...
 #   custom:      any model name your OpenAI-compatible endpoint accepts
@@ -210,6 +210,9 @@ ENGRAPHIS_LLM_MODEL=gpt-4o-mini
 # ENGRAPHIS_LLM_BASE_URL=https://openrouter.ai/api/v1
 # Optional: extra headers (JSON string) for custom providers.
 # ENGRAPHIS_LLM_EXTRA_HEADERS={"HTTP-Referer":"https://myapp.com","X-Title":"engraphis"}
+# Optional: reasoning effort (low|medium|high|xhigh|max) for Claude models that think by
+# default (Opus 5+, Sonnet 5+, Fable). Default medium; ignored by other providers and models.
+# ENGRAPHIS_LLM_EFFORT=medium
 
 # ── Hosted Pro / Team customer client ───────────────────────────────────────
 # Cloud Sync, Analytics, Automation, Auto Dreaming, Auto Consolidation, and Team

BENCHMARKS.md

@@ -94,14 +94,14 @@ interpretation and do not count as additional benchmark-quality gains.
 ### Public numeric evidence registry
 
 Every exact public aggregate retained below comes from the checked-in, public-safe
-[`offline-fixtures-v117.json`](docs/benchmark-evidence/offline-fixtures-v117.json) artifact. Its
+[`offline-fixtures-v125.json`](docs/benchmark-evidence/offline-fixtures-v125.json) artifact. Its
 SHA-256 is
-`c64eadffbc7f87938e0e822b7aab33dcfece6fffa5175932ac208465d36df192`, also recorded in the
+`1f74971d6213a188b31cf58f6ff6132a487da36d22455292b028dadc202a8feb`, also recorded in the
 adjacent `.sha256` file. The artifact contains no raw questions, answers, prompts, customer data,
 or per-record content fingerprints.
 
 The fixture-suite digest is
-`77b140068cd036330b421a0d7668847d196f1bbeaa36c965d68943457002e6eb`. The artifact defines
+`42cb269b867a1e3210d0da77d6a040abc974cd44317b91ad10a2d483d07c1586`. The artifact defines
 the digest algorithm and records the SHA-256 of every suite and dataset file. Each evidence ID
 also binds its exact command through `sha256(UTF-8 exact command)`:
 
@@ -123,10 +123,10 @@ Historical LoCoMo, graph, handoff, consolidation, and security figures remain pr
 source artifacts but are omitted from the current chart until each has a matching immutable,
 public-safe artifact. The chart labels coding outcomes, external datasets, and operational
 capacity as pending evaluation tracks rather than implying scores. Regenerate it with
-`python scripts/render_benchmark_report.py --report docs/benchmark-evidence/offline-fixtures-v117.json --output docs/images/context-efficiency.svg` after selecting the report to publish.
+`python scripts/render_benchmark_report.py --report docs/benchmark-evidence/offline-fixtures-v125.json --output docs/images/context-efficiency.svg` after selecting the report to publish.
 
 The companion examples are also generated from that artifact with
-`python -m scripts.render_benchmark_examples --report docs/benchmark-evidence/offline-fixtures-v117.json --output docs/images/evidence-backed-agent-examples.svg`.
+`python -m scripts.render_benchmark_examples --report docs/benchmark-evidence/offline-fixtures-v125.json --output docs/images/evidence-backed-agent-examples.svg`.
 The historical-to-executable mapping is in
 [`docs/BENCHMARK_CHANGE_COVERAGE.md`](docs/BENCHMARK_CHANGE_COVERAGE.md).
 

CHANGELOG.md

@@ -5,6 +5,42 @@ All notable changes to Engraphis are documented here. Format loosely follows
 
 ## [Unreleased]
 
+- Classify local `engraphis_decide` command advice conservatively. Chained, piped,
+  substituted or redirected commands, and options that run programs or write files
+  (`rg --pre`, `pytest --basetemp`), are never labeled read-only. Recursive deletes in any
+  flag order, raw device writes, history-rewriting or work-discarding Git operations, SQL
+  and infrastructure teardown, downloaded-script execution, exfiltrating pipes and uploads,
+  and well-known credential files are labeled destructive or leaking. Screening is bounded
+  so adversarial input stays cheap.
+- Match whole words, ignore zero counts and treat negated success as failure in local
+  completion checks; contrasts ("not only passed") and negated failures ("did not fail") are
+  not failures. Local support and contradiction checks compare content words rather than
+  shared stopwords, and support keeps one-character terms such as "C". Supersession requires
+  a shared subject plus a replacement cue the existing fact lacks, or a negation (including
+  "cannot") whose ruled-out clause the other fact asserts ("does not use port 80" opposes
+  "uses port 80", not "uses port 443"); between terse facts one shared word is the subject,
+  so "No SQLite" opposes "Use SQLite". Reinforcement requires matching cues and one fact
+  containing the other's content words, so conflicting values defer. Any number of Git global
+  options and clustered short flags no longer bypass destructive-command checks, and
+  path-specific or pathspec-file checkouts, worktree restores and force-created branch resets
+  count as discarding work. Credential paths match either path separator.
+- Fixed Anthropic connections for current Claude models. Opus 4.7 and later, Sonnet 5 and
+  later, and Fable no longer receive `temperature`, which they reject. Replies are read from
+  text blocks, so a leading thinking block no longer fails with "Unexpected Anthropic response
+  format". Models that think by default get `ENGRAPHIS_LLM_EFFORT` (default `medium`) and at
+  least 4096 output tokens so reasoning cannot crowd out the reply.
+- Replaced the retired `claude-3-5-sonnet-20241022` Anthropic default in the dashboard picker,
+  the API defaults, `.env.example`, and the provider guide with `claude-sonnet-5-5`.
+- Documented how to back the experimental Jev decision adapter with Claude: pin an exact model
+  id, keep fallback disabled, and avoid sampling parameters and forced tool choice.
+- Recognized `claude-mythos-preview` and other unversioned Fable/Mythos ids, so they no longer
+  receive `temperature` (an HTTP 400) and get the thinking-model output headroom.
+- A remote `verify_completion` probability between the certainty bound and the 0.85 completion
+  bar is now `uncertain` with a null `is_complete`, instead of a decisive failure.
+- Local completion checks recognize `N passing` (Mocha/Jest style), `not passing`, `0 passing`
+  and `errors: none`.
+- Refreshed the public offline fixtures and source bindings in immutable v125 evidence.
+
 ## [1.7.9] - 2026-09-29
 
 - Added the saved-session managed Jev transport and Classic/Smart MCP decision route
@@ -53,7 +89,8 @@ All notable changes to Engraphis are documented here. Format loosely follows
   session or repo mapping, report the resolved destination, and reject session mismatches.
 - Command Code's SessionStart hook now uses the nearest Git root's repo name, honors saved
   workspace mappings unless explicitly overridden, and labels recalled context with the
-  server's resolved workspace.
+  server's resolved workspace. Save a mapping to keep recalling memories stored under the
+  earlier folder-named workspace default.
 - Added a previewed selective move workflow for organizing mixed workspaces while retaining
   source history and enforcing move eligibility and workspace access.
 - Hardened the experimental Cloud decision client with validated destinations,

README.md

@@ -84,9 +84,9 @@ neither is an end-to-end question-answer score. Coding outcomes, external datase
 operational capacity remain separate pending evaluation tracks until their artifacts are selected.
 
 These values are evidence IDs `offline-chunking` and `offline-performance` in
-[`offline-fixtures-v117.json`](https://github.com/Coding-Dev-Tools/engraphis/blob/main/docs/benchmark-evidence/offline-fixtures-v117.json),
+[`offline-fixtures-v125.json`](https://github.com/Coding-Dev-Tools/engraphis/blob/main/docs/benchmark-evidence/offline-fixtures-v125.json),
 SHA-256
-`c64eadffbc7f87938e0e822b7aab33dcfece6fffa5175932ac208465d36df192`.
+`1f74971d6213a188b31cf58f6ff6132a487da36d22455292b028dadc202a8feb`.
 [`BENCHMARKS.md`](https://github.com/Coding-Dev-Tools/engraphis/blob/main/BENCHMARKS.md#public-numeric-evidence-registry)
 records the matching suite digest, exact commands, and per-command config digests. The offline
 fixture registry intentionally excludes external, model-dependent, consolidation, productivity,
@@ -442,6 +442,8 @@ open on timeout and is installed via `python scripts/install_cc_hook.py`.
 The hook sends the nearest Git root's name as `repo` and lets the server apply a saved workspace
 mapping. Set `ENGRAPHIS_HOOK_WORKSPACE` only for an explicit override; a previous `default`
 override must be cleared to use the mapping. Its context header shows the resolved workspace.
+Earlier versions defaulted to a workspace named after the project folder; save that mapping to
+keep recalling those memories at session start.
 
 ### prime-agent fleet
 
@@ -848,6 +850,7 @@ file. It never searches the working directory for `.env`, and explicit process v
 | `ENGRAPHIS_LLM_MODEL` | `gpt-4o-mini` | Model name (provider-specific) |
 | `ENGRAPHIS_LLM_API_KEY` | Not set | API key for chat/synthesis, `llm` / `llm_structured` extraction, and structured consolidation |
 | `ENGRAPHIS_LLM_BASE_URL` | Not set | Base URL for openrouter / custom OpenAI-compatible endpoints |
+| `ENGRAPHIS_LLM_EFFORT` | `medium` | Reasoning effort (`low \| medium \| high \| xhigh \| max`) for Claude models that think by default (Opus 5+, Sonnet 5+, Fable); ignored by other providers and models |
 | `ENGRAPHIS_DECISION_BACKEND` | `none` | `none` or `local` keeps advisory decisions local; `managed` uses the saved Cloud session and included allowance; `auto` selects managed when configured and never switches to BYOK; explicit `byok` uses a personal TypeSafe key. Legacy `typesafe`, `jev`, and `system1` mean BYOK. Remote calls also require per-call consent. |
 | `ENGRAPHIS_DECISION_MODEL` | `jev-1.13.0` | Pinned model accepted by the Jev transport; other model identifiers are rejected. |
 | `TYPESAFE_API_KEY` | Not set | Personal credential for explicit BYOK decisions; `JEV_API_KEY` is a fallback alias. Managed decisions use the saved Cloud session instead. |

docs/LLM_PROVIDERS.md

@@ -54,6 +54,7 @@ Every LLM setup uses these variables:
 | `ENGRAPHIS_LLM_API_KEY` | Credential for the provider. It is never returned by the dashboard. |
 | `ENGRAPHIS_LLM_BASE_URL` | Needed only to override a default or configure a compatible endpoint. |
 | `ENGRAPHIS_LLM_EXTRA_HEADERS` | Optional JSON object of headers required by a compatible endpoint. |
+| `ENGRAPHIS_LLM_EFFORT` | Reasoning effort for Claude models that think by default: `low`, `medium` (default), `high`, `xhigh`, or `max`. Ignored elsewhere. |
 
 The sample names below are Engraphis runtime defaults, not provider recommendations. Replace them
 when your account or deployment uses a different model.
@@ -91,13 +92,30 @@ it as `custom`; the native mode applies Anthropic's required request shape and h
 
 ```dotenv
 ENGRAPHIS_LLM_PROVIDER=anthropic
-ENGRAPHIS_LLM_MODEL=claude-3-5-sonnet-20241022
+ENGRAPHIS_LLM_MODEL=claude-sonnet-5-5
 ENGRAPHIS_LLM_API_KEY=<anthropic-api-key>
 ```
 
 Leave `ENGRAPHIS_LLM_BASE_URL` unset for the public API. For model and credential details, see
 the [Anthropic API documentation](https://docs.anthropic.com/).
 
+Choosing a model:
+
+- `claude-sonnet-5-5` is the default. Extraction, consolidation summaries, and grounded synthesis
+  are bounded tasks, and it costs half as much per token as Opus.
+- `claude-opus-5-5` suits the hardest consolidation and conflict-review work. Compare it against
+  the Sonnet default on your own data before paying for it, because the two are close on everyday
+  tasks.
+- Retired ids such as `claude-3-5-sonnet-20241022` and `claude-3-5-haiku-20241022` are rejected by
+  the API; the connection test reports them as an HTTP 404.
+
+Opus 4.7 and later, Sonnet 5 and later, Fable and Mythos (including `claude-mythos-preview`)
+reject `temperature` and similar sampling parameters, so Engraphis omits them for those models.
+Opus 5 and later, Sonnet 5 and later, Fable and Mythos also think before answering by default. Engraphis sends `ENGRAPHIS_LLM_EFFORT` (`low`,
+`medium`, `high`, `xhigh`, or `max`; default `medium`) for those models and keeps at least 4096
+output tokens available so hidden reasoning cannot crowd out the reply. Other providers and older
+Claude models ignore the setting.
+
 ## Google Gemini
 
 Google Gemini uses the native `google` mode and the Gemini `generateContent` API. The native mode

docs/MCP_TOOLS.md

@@ -190,6 +190,10 @@ and `query`; `verify_completion` needs `state` and `goal`, with optional `recent
 before backend lookup, with unknown/null conclusions and no remote allowance consumed.
 Command decisions always return `allow_auto=false` and `escalate_to_user=true`, including
 successful remote answers. Provider probability and category are advice, not shell authorization.
+Local command labels are coarse: only one simple inspection command, without chaining, pipes,
+substitution or file redirection, is `read_only`. Recognized destructive, history-rewriting,
+exfiltrating or credential-file commands are `destructive_or_leak`; anything else, including a
+command too long to screen completely, is `state_change`.
 
 The classic recall, grounded, and answer tools (`engraphis_recall`,
 `engraphis_recall_grounded`, and the `engraphis_answer` alias) accept `planning="off"|"auto"`,

docs/WORKSPACE_ORGANIZATION.md

@@ -111,6 +111,11 @@ A nonblank `ENGRAPHIS_HOOK_WORKSPACE` is an explicit override. Clear a previous
 override to use project mappings. The recalled-context header names the workspace returned by
 the server. The hook remains silent on errors or empty recall results.
 
+Earlier hook versions used a workspace named after the project folder when no override was set.
+Without a saved mapping, the hook now starts in `default` instead, so those memories stop
+appearing at session start. To keep using them, save the mapping once, for example
+`engraphis_set_workspace_routing(workspace="website", repo="website")`, or move the memories.
+
 ## Organize existing memories
 
 Routing changes future writes. Existing memories stay where they are until explicitly moved.