Add Qwen inference gateway and current Codex transcript support

.env.example

@@ -7,3 +7,12 @@ VOYAGE_API_KEY=
 # Optional only for a portable/custom Cursor profile. The standard macOS/Linux
 # database path is detected automatically.
 SESSION_RECALL_CURSOR_DB=
+
+# Optional Qwen inference gateway; see .project-docs/processes/inference-api-setup.md.
+# SESSION_RECALL_EMBED=inference-api
+# SESSION_RECALL_EMBED_BASE_URL=https://inference.example/v1
+# SESSION_RECALL_EMBED_DIM=4096
+# SESSION_RECALL_INFERENCE_MAX_TOKENS=8192
+# SESSION_RECALL_EMBED_REVISION=deployment-revision
+# SESSION_RECALL_DB_PATH=/path/to/separate/index-qwen.db
+INFERENCE_API_KEY=

.project-docs/agent-context.md

@@ -0,0 +1,33 @@
+<!-- Synchronized project guidance. Canonical editable path: .project-docs/agent-context.md. -->
+
+# Project guidance
+
+Project knowledge lives under `.project-docs/`.
+
+Before making a non-trivial project-specific claim:
+
+1. read `.project-docs/manifest.yaml`;
+2. follow `.project-docs/navigation.md`;
+3. search the routed locations and read the canonical record plus material
+   related links;
+4. check status, freshness, and sources.
+
+Use `services/` for current systems, `processes/` for exact actions,
+`decisions/` for rationale, `reactions/` for conditional first actions,
+`bugs/` for known failures, and `timeline/` for chronology.
+
+Do not treat observations, stale claims, conflicts, recommendations, or
+unknowns as confirmed facts. Never store, quote, partially reproduce,
+transform, or echo secret values.
+
+Change project guidance only through the `project-documentation` workflow at
+the canonical editable source, `.project-docs/agent-context.md`. Do not edit
+`AGENTS.md` or `CLAUDE.md` directly; both generated targets are byte-identical
+to the canonical source.
+
+After every documentation mutation, run
+`python3 <skill-directory>/scripts/validate_project_documentation.py
+--project-root <project-root> --tracked-documentation`. Add `--include` for a
+new untracked documentation file. If canonical guidance changed, also run the
+guidance synchronization script with `--diff`, `--write`, and `--check`.
+Documentation work is incomplete until validation and synchronization pass.

.project-docs/manifest.yaml

@@ -0,0 +1,31 @@
+schema_version: 1
+entrypoints:
+  project: project.md
+  navigation: navigation.md
+routes:
+  current_state:
+    - project.md
+    - services/
+  how_to:
+    - processes/
+  why:
+    - decisions/
+  incident_first_action:
+    - reactions/
+    - processes/
+    - services/
+  known_problem:
+    - bugs/
+  history:
+    - timeline/
+    - changelog/
+  unresolved:
+    - open-questions.md
+    - observations/
+search_fields:
+  - id
+  - title
+  - summary
+  - tags
+  - canonical_for
+  - related

.project-docs/navigation.md

@@ -0,0 +1,28 @@
+# Documentation navigation
+
+## Search order
+
+1. Classify the question using `manifest.yaml` routes.
+2. Search only routed paths with task terms, identifiers, tags, and synonyms.
+3. Read the best canonical match.
+4. Follow only material `related` links.
+5. Check status, freshness, and sources before using a claim.
+
+Example:
+
+```bash
+rg -n -i 'trace|tracing|observability' \
+  .project-docs/services .project-docs/processes .project-docs/decisions
+```
+
+## Stopping rules
+
+- No canonical match means unknown.
+- Observations remain unconfirmed.
+- Stale records require re-verification.
+- Conflicts preserve every sourced version.
+- Recommendations are not current behavior.
+- Missing sources invalidate confirmed claims.
+
+Record missing knowledge in `open-questions.md`; do not invent a project
+default.

.project-docs/open-questions.md

@@ -0,0 +1,6 @@
+# Open questions
+
+- Gateway deployments must supply their own endpoint, credentials, dimensions,
+  context limit, and embedding revision. These are not repository defaults.
+- The gateway registry does not provide a required immutable embedding revision;
+  operators must update the configured revision when the backend changes.

.project-docs/processes/inference-api-setup.md

@@ -0,0 +1,66 @@
+---
+id: inference-api-setup
+type: process
+title: Connect a Qwen inference gateway
+summary: Configure the gateway and build a separate index before switching retrieval.
+status: confirmed
+tags: [qwen, setup, migration, backup]
+canonical_for: [inference-api-setup]
+verified_at: 2026-09-14
+sources:
+  - type: repository
+    reference: src/session_recall/config.py
+    confirmed_at: 2026-09-14
+  - type: repository
+    reference: src/session_recall/inference.py
+    confirmed_at: 2026-09-14
+  - type: repository
+    reference: src/session_recall/cli.py
+    confirmed_at: 2026-09-14
+related: [../project.md, ../services/inference-api.md]
+---
+
+# Connect a Qwen inference gateway
+
+Check the [required gateway contract](../services/inference-api.md) first.
+Obtain the endpoint, embedding dimension and context limit from the deployment
+owner or authenticated model registry. Supply `INFERENCE_API_KEY` through your
+secret manager or environment; never commit its value.
+
+Example configuration for a gateway with 4096-dimensional vectors and an
+8192-token context window:
+
+```sh
+export SESSION_RECALL_EMBED=inference-api
+export SESSION_RECALL_EMBED_BASE_URL=https://inference.example/v1
+export SESSION_RECALL_EMBED_DIM=4096
+export SESSION_RECALL_INFERENCE_MAX_TOKENS=8192
+export SESSION_RECALL_EMBED_REVISION=deployment-revision
+export SESSION_RECALL_DB_PATH="$HOME/.local/share/session-recall/index-qwen.db"
+session-recall index
+session-recall health
+session-recall search "why did we choose"
+```
+
+The preset uses model aliases `embedder` and `reranker`. Override them with
+`SESSION_RECALL_EMBED_MODEL` and `SESSION_RECALL_RERANK_MODEL` if your gateway
+uses different public names. Existing provider overrides also take precedence
+over the preset, so remove or update stale overrides when migrating.
+
+Before migration, preserve the old provider settings and make a consistent
+SQLite backup of the old index. Build the new embedding space in a separate
+file with `SESSION_RECALL_DB_PATH`; do not mix Voyage and Qwen vectors.
+Review indexing errors, corpus coverage and health before switching clients.
+An unavailable source database cannot supply new history; retain its old
+index until its historical records have been accounted for.
+
+Configure CLI invocations, background indexers, and MCP launchers with the same
+provider settings and database path. Restart existing MCP processes after
+switching, because they retain their configuration and open database.
+To roll back, restore the previous launcher configuration and use the old
+index with its original embedding provider.
+
+When the backend behind the public alias changes incompatibly, update
+`SESSION_RECALL_EMBED_REVISION` and rebuild. A stable model alias alone does
+not identify an immutable embedding space. Session Recall does not auto-load
+`.env` files; export these settings or load them through your launcher.

.project-docs/project.md

@@ -0,0 +1,25 @@
+---
+id: session-recall-project
+type: project
+title: Session Recall
+summary: Semantic retrieval over local Claude Code, Codex, and Cursor history.
+status: confirmed
+tags: [recall, indexing, embeddings]
+canonical_for: [project-overview]
+verified_at: 2026-09-14
+sources:
+  - type: repository
+    reference: README.md
+    confirmed_at: 2026-09-14
+related: [services/inference-api.md, processes/inference-api-setup.md]
+---
+
+# Session Recall
+
+Session Recall extracts conversation text, stores text and vectors in SQLite,
+and exposes retrieval through its CLI and MCP server. Provider selection and
+index identity are configured outside the repository.
+
+See the [Inference API contract](services/inference-api.md) and
+[connection procedure](processes/inference-api-setup.md) for hosted Qwen gateways.
+The root README covers existing providers and the remaining product workflows.

.project-docs/services/inference-api.md

@@ -0,0 +1,56 @@
+---
+id: inference-api-provider
+type: service
+title: Qwen inference gateway provider
+summary: Embeddings and reranking through an authenticated capability-based HTTP API.
+status: confirmed
+tags: [qwen, inference, embeddings, reranker, codex]
+canonical_for: [inference-api-provider]
+verified_at: 2026-09-14
+sources:
+  - type: repository
+    reference: src/session_recall/inference.py
+    confirmed_at: 2026-09-14
+  - type: repository
+    reference: src/session_recall/config.py
+    confirmed_at: 2026-09-14
+  - type: repository
+    reference: tests/test_inference.py
+    confirmed_at: 2026-09-14
+related: [../project.md, ../processes/inference-api-setup.md]
+---
+
+# Inference API provider
+
+The `inference-api` preset selects the public `embedder` and `reranker` model
+aliases. It defaults to 4096 embedding dimensions; deployments can override
+that value. The base URL and credential have no default.
+
+The client sends `POST embeddings` beneath the configured `/v1` base URL,
+with `input_type=document` for indexing and `input_type=query` for retrieval.
+It sends batches of up to 64 strings and requests base64 float32 vectors.
+Response indices determine ordering. Both float arrays and base64 responses
+are validated for dimension and finite values before storage.
+
+Reranking uses `POST rerank` with `query`, `documents`, and `top_n`. Results
+must contain distinct valid document indices and finite `relevance_score`
+values. The client returns scores in descending order.
+
+Both endpoints receive `truncate_prompt_tokens`, defaulting to 8192. This
+limits model input; the full extracted text remains in the local index.
+The gateway must support these request fields, including the query/document
+instruction distinction; a generic embeddings-only API is not sufficient.
+
+The client uses `INFERENCE_API_KEY`, verifies HTTPS for remote endpoints,
+does not follow redirects, and ignores proxy environment variables. Local
+loopback HTTP is supported. Requests use a 120-second timeout with a
+10-second connection timeout and up to three attempts for transport errors,
+429, and selected temporary server errors. HTTP error messages omit response
+bodies so echoed transcripts do not enter indexing logs.
+
+Index fingerprints include the provider, model alias, dimension, endpoint,
+operator-supplied revision, token limit, and query/document preprocessing
+version. Changing any of these requires compatible reindexing. Encoding and
+batch size do not change the embedding space.
+
+See the [setup procedure](../processes/inference-api-setup.md).