core: index-only counts and search positions ...

The sidebar's per-session counts and the search page's match positions ran
over the message heap: rows are wide (content, thinking) and the session
index has no is_display, so every counted row cost a heap fetch. A covering
index on (session_id, sequence_number) INCLUDE (is_display) turns both into
index-only scans.

The match query now numbers positions over that projection and reads content
only for the rows the trigram index returns, instead of numbering every
display row together with its content. On prod (1.3k sessions, 77k messages)
a common term across the largest sessions went from ~1.7 s to ~0.1 s, and the
summary aggregate from ~0.5 s to the index.

core: session lists read aggregates, search matches from SQL ...

GET /sessions without paging hydrated every hot message of every listed
session just to return a count and a preview: 313 ms locally, and the whole
hot history allocated. `list_all_summaries` returns the same rows in the same
order with the count and snippet from one aggregate — 30 ms.

`search_list` now returns summaries like `list_page`; "where does the term
appear" comes from `search_matches`, one aggregate that reproduces the display
indices clients scroll to (verified equal to the old Python scan on 105
sessions). The API search path drops from ~500 ms to ~150 ms, most of which is
the ILIKE scan itself — see docs/sessions.md for the pg_trgm note.

The stale-session sweep on POST /sessions takes its candidates from summaries
and loads a history only for a session that actually needs extraction.

protocol, clients: acks, queue notices and reconnect without duplicates ...

- client_id is echoed in message_accepted / message_queued; queue overflow
  broadcasts message_dropped to every socket watching the session
- stream_end / stream_stopped carry message_index, the global display rank
  that also stamps history rows
- webclient: cheap session_sync via /sessions/{id}/meta, the ack deadline is
  armed at the write and re-armed while a run is active, fatal 4003/4004 stop
  the reconnect loop instead of retrying forever
- TUI: WsBridge posts WsResumed / MessageWritten / SessionUnavailable, the app
  drops the half-written turn on reconnect, shows queued and not-delivered
  markers on the bubble, and reloads history only when the meta token changes
- dist rebuilt

core, webclient: cheap session lists, changed-row saves, smaller bundle ...

The sidebar asked for a count and a 60-char snippet and got every hot
message of every listed session: 2,097 rows / 4.4 MB for 507 sessions,
plus a full Pydantic hydration. list_page now aggregates counts and the
last displayed line in SQL (4.0 ms, 30 rows); get_meta folds its COUNT
into the same round trip.

save() runs after every streaming step and rewrote the whole hot
history each time — 1,016 UPDATEs, 46.8 ms, per call, growing with the
session. A fingerprint of the mutable columns lets it touch only the
rows that changed, which is normally the tail.

HTTP: content-hashed /assets are immutable for a year, gzip drops to
level 6 (same ratio within a couple of percent on JSON, half the CPU),
and session search gets pg_trgm GIN indexes when the extension exists.

Client: one shared highlight.js on lib/common plus the languages we
actually render (openscad included) instead of the full ~190-language
entry point, and the unused @tanstack/vue-virtual is gone.
Bundle 1566 -> 760 kB (gzip 495 -> 238 kB).

tools: split filesystem into read, write and file_ai ...

- filesystem_read: read, list, find, find_up, grep, info, exists, diff
- filesystem_write: write, append, edit, edit_lines, insert, mkdir, move, copy, delete
- file_ai: ask (was query), edit (was smart_edit)
- paged numbered reads (500 lines / 50 KB), raw copy mode, binary refusal
- nearest-match and numbered-copy hints on failed edits, context after edit
- guards: empty overwrite, non-empty delete, move/copy overwrite
- cross-tool redirects instead of silent wrong-action errors
- legacy 'filesystem' stays for archived sessions (compression, renderers, webclient)
- profiles, prompts, docs and manuals updated; TUI and webclient render the new tools

core: anchor restricted-profile boundaries in subagent prompts ...

Sub-agents of non-admin profiles replaced the parent system prompt with
their own worker prompt, so the "one directory, no network, no secrets"
rules never reached them. build_subagent_system_prompt now appends a
[Boundaries] block for every non-admin profile, whatever prompt was
chosen.

Align the profile prompts with the actual tool surface alongside it:
allow 3D sub-agents to draft geometry but not publish or decide, publish
a preview PNG with the STL in both 3D profiles, document the assistant's
remaining tools, canonicalise MCP names, and fix the developer/navi_code
spawn ending, the server_admin field list, the discuss memory_search
guidance, and the navi-3d/navi-web/gnexus-book instructions.

core: split the agent and context builder into packages ...

Agent becomes one class assembled from mixins: runtime (turn lifecycle),
streaming (the turn loop), tools (tool execution), compression (context
gates) and context (tool lists, backend lookup, window). The module-level
helpers and sentinels move to helpers.py.

ContextBuilder splits into helpers (pure truncation/freshness/prompt
blocks) and builder (the class). Both old modules become packages with
the same import surface, so navi.core.agent.Agent and
navi.core.context_builder.ContextBuilder keep working.

Method bodies are unchanged; tests now patch settings/log/should_compress
on the mixin that actually reads them.

core: split the Postgres session store into a package ...

`pg_session_store.py` mixed three layers: schema DDL and migrations, the
row <-> `Message` mapping with its archive-aware hydration, and the store
itself. Split them into `navi/core/sessions/`:

* `ddl.py` — table/index definitions and the one-shot migrations;
* `rows.py` — serialization, dangling-tool-call repair, session hydration;
* `store.py` — `PgSessionStore` and the pending-sweep loop.

`navi/core/pg_session_store.py` stays as a compatibility shim re-exporting
both the public surface and the private helpers the tests use. No behavior
change: every top-level function and every store method is preserved.

core: split the compressor into a package ...

`navi/core/compressor.py` was 1355 lines covering three different jobs:
deciding what a pass summarizes, building and calling the summarizer, and
the retry/truncate/persist wrapper around it. Move each to its own module:

* `compression/partition.py` — pure partitioning and the single
  `_plan_compression` decision shared with `would_compress`;
* `compression/summarize.py` — prompt formatting, the KEYS line, the
  meta-summary and `compress_context`;
* `compression/compressor.py` — `ContextCompressor` and its fallbacks.

`navi/core/compressor.py` stays as a compatibility shim that re-exports the
public surface, so imports elsewhere are untouched. Because each module now
binds its own `settings`, the tests that override settings patch all of them
(and `_meta_summarize`/`compress_context` are patched where they are read,
not on the shim). Adds a regression test for the hard-truncate trim inside a
single oversized turn, which the moved `_split_turn_tail` import silently
broke (ruff F821, no test hit that branch).

memory: fix review findings in unified search ...

- index each compression-dropped turn once: messages and context share
  objects, and earlier drops stay in the display history
- label raw excerpts against facts even when recall asks for scope=raw
- top up short vector results from the lexical index (NULL embeddings)
- count whole content words only and cap query terms at 12
- carry source_session_id on fact hits
- cap injected_chunk_ids at 100 entries
- keep a session delete successful when chunk cleanup fails
- stop logging query text in the memory paths
- stop fabricating similarity scores for recent chunks
- consult session_messages_archive in latest_summary_at

memory: one search across facts, raw history and sketches ...

Compression removes messages from the model's window, not from the
conversation: every compression path now writes the turns it demotes into
`memory_chunks` — verbatim, append-only, with the turn range, timestamps and
breadcrumb keys (the summarizer's `KEYS:` line plus a deterministic extractor,
which also covers the hard-truncate fallback that never sees a summarizer).

`memory_search` is the single read path over the three stores: curated facts
first, then raw excerpts by relevance, then the current session's sketches.
Every hit carries provenance (source, session, turn range, timestamp), and a
raw excerpt that a newer curated fact overrides is labelled outdated rather
than hidden — "we changed our mind, the old value was X" is often the
question. The agent also gets up to two relevant excerpts injected per turn
once the context holds a summary, deduped across turns.

Fixes found by running the pilot against a real database:
- `search_facts` returns ISO-string timestamps while chunks return datetimes;
  the freshness comparison raised TypeError and took the whole search down.
- an existing `memory_chunks` table created without pgvector never gained the
  `embedding` column (CREATE TABLE IF NOT EXISTS does not alter), and the
  INSERT failed on it; the DDL now adds the column, and a database without
  pgvector writes text-only rows found through FTS.

context: price images by area and detect backend-truncated prompts ...

Every window decision priced an inline image at a flat 500 tokens, while a
1024x1024 screenshot costs ~1026 on qwen3.6 and a 2048x2048 one ~4098. The
undercount kept sessions below the 90% trigger; Ollama then silently truncated
the prompt to about half the window (dropping the system prompt), and the
truncated prompt_eval_count was cached as the session baseline, pinning the
estimate at ~50% for good - the session never compressed while the model was
already told the context was full.

- image_tokens.py prices each image from its header (one token per 32x32 px,
  downscaled to <=2048^2, clamped to 256..4096); unreadable headers fall back
  to CONTEXT_IMAGE_TOKENS. Used by both estimators, the delta and
  check_context_size; the per-turn image budget fits images by their own cost.
- A prompt reported far below what was sent is a truncation: it is not cached
  as the baseline, and the post-turn worker compresses regardless of the
  threshold. The session keeps the honest token count.
- The pre-turn gate prices the incoming message and its images, which are not
  in session.context yet.
- The summarizer receives images under a window-scaled budget, and prompts
  carrying images stay out of the calibration fit (their bytes are not in
  chars and would skew it).
- Oversized user messages are capped for the LLM view (at least num_ctx/3),
  so a paste bigger than the window no longer dead-ends a session.
- The in-flight partition also trims by token budget: a gate that fires can
  now actually shrink a turn made of a few huge messages.

sessions: share a chat as a public read-only link ...

A session's owner can hand out one capability link that opens the display
history without an account: no input, no chat list, no artifacts panel.
The token lives in the URL hash, so it never reaches the server in a request
line or a Referer header.

The link is permanent by decision - no expiry, no revoke. Deleting the chat
removes the share row (FK cascade) and closes it, which is the only switch.

Backend: session_shares table (unique token, FK cascade), POST
/sessions/{id}/share issuing the session's one link (owner or admin; a
read-all permission may read another user's chat but not publish it), and
public GET /shared/{token}[/messages] serving display history only, with
session and owner ids left out of the payload.

Webclient: #share=<token> renders a standalone page with the owner's cards,
rating off, interactive UI components filtered out and scroll-up paging; a
share button in the chat header opens a dialog that creates the link.

Also fixes ChatHeader's rename, which called an undefined `api` and failed
silently.

navi: overlay the hive config onto the instance's own ...

An instance that knows a hive with the swarm key starts from the hive
document: layers that apply to it are merged over its local config, fields
the hive does not mention stay local, and an empty answer means 'run exactly
your own config'. Settings, profiles, LLM routes and MCP servers are covered;
database URL, ports, auth switches and the channel's own settings can never be
overridden.

Startup fetches with a short timeout, falls back to the last validated overlay
on disk and refuses to crash-loop on a document that killed the previous start.
A live refresh applies secrets and timeouts in place (the frozen singleton is
written through one deliberate escape hatch), rebuilds profiles from their
local base and reports everything else as restart_required. The hive reaches
the instance by push, by the announce reply and by a retry sweep; every apply
is acked back, which is what the hive's status view reads.

profiles: teach the agent to scout profiles and tools before switching ...

Persona gains a reconnaissance order (list_profiles, then list_tools for the
real toolset, then tool_manual) and the switch-vs-delegate rule: switch when
the task itself is another profile's work, delegate a side piece with
spawn_agent. The profiles block is built in one place and shared with the
admin prompt debug view, and the stale 'next user message' claim is fixed.

list_tools applies the admin-only role gate to an explicitly named profile,
the same one list_profiles uses, so it cannot enumerate hidden profiles.

prompt: give the methodologies catalog its rules in the system prompt

protocol: run a protocol as a committed plan, with a log of how it ended ...

A protocol was only text until now. The new `protocol` tool starts one: its
steps are spliced into the session's todo plan after the agent's current
position (nothing already done is reordered), a row appears in `protocol_runs`
with a snapshot of the version, steps and post-checks, and the run stays
`running` until it is closed.

`finish` refuses while any step of the run is open, refuses outright when one
failed, and requires a report — it is the run's record for the user. `cancel`
and `fail` require a reason and mark the rest of the run's steps `skipped`, so
the plan does not keep work that was never done. One run per session is active
(a partial unique index enforces it).

The commitment also has to survive the context: every model turn while a run is
open carries a system reminder with the protocol, its steps and the way to close
it, and the final-turn intercept now fires for an open run even when the todo
shows no open steps, so a bare-text turn cannot leave one running.

`protocol` is wired into assistant, developer, discuss and server_admin next to
`skills`; sub-agents get neither, because a protocol belongs to the session's
plan. The run log has no UI yet — `protocol(op='status')` is the current view.

skills: catalog of skills and protocols with a settings tab ...

One registry for reusable methodologies: markdown with YAML front matter,
stored in PostgreSQL alongside a full edit history. A row is either a
`skill` (advice, adapt as needed) or a `protocol` (a committed order of
actions), and a protocol is refused on save unless it declares its steps,
post-checks and the three escape hatches — on_deviation, on_blocked,
on_already_done — because an honest run needs a way out when reality
differs from the plan.

- `SkillStore`: lazy DDL, uniqueness on (owner_id, kind, name), version
  bump with an appended `skill_versions` row, `content_hash` as the only
  hive-sync hook. Personal rows carry an owner; shared rows have an empty
  one and are writable by an admin.
- REST `/skills` (list/get/create/update/delete/versions) and the agent
  tool `skills` (search/load) that renders a compact document — numbered
  steps and escape hatches for a protocol, the instructions for a skill.
- Settings → Скилы и протоколы: catalog with search and kind/area
  filters, a document editor that shows the server's validation error
  verbatim, create from a template, delete, edit history, and the
  personal/shared switch for an admin.

Manuals stay where they are: tool knowledge, separate from this catalog.

todo: give plans their own table and recover them after a crash ...

Todo already survived restarts inside the generic session_store KV row, but
the plan there was opaque JSON with no timestamp: after a server crash the
agent resumed a step marked in_progress with no way to tell whether it had
ever been done.

session_todos (user_id, session_id, tasks JSONB, revision, updated_at) now
holds the plan. Existing rows are migrated on first use and the KV source is
deleted in the same step, so clearing a plan is not undone by the next
process start.

Each step gains provenance fields (source, step_id, protocol_run_id) for the
protocol work, and insert_steps() splices a block after the step the agent
has already reached instead of making it re-issue the whole plan.

get_progress_message() now flags a plan whose in_progress step predates this
process once per (session, revision): re-check the real result before
marking it done.

The todo op "clear" now uses the same user scope as load/save; it used to
prefer ctx.user_id and could delete a different row than the one written.

subagents: a stopped run closes the tool it had opened ...

A tool card is drawn from tool_started and only its tool_call closes it.
When the user stopped a run while a sub-agent was mid-tool, the sub-agent
was cancelled, the parent dropped its sink, and the matching tool_call was
never sent — the card spun until the page was reloaded.

The sub-agent now reports its in-flight tool from the CancelledError
handler, and both batch paths (sequential and parallel) drain the sink
after cancelling instead of discarding it, so that event reaches the
client ahead of the parent's synthetic "stopped by the user" result.

The webclient settles anything still pending at stream_end,
stream_stopped and error as a safety net for a lost frame or a dying
server, with a visible "interrupted" note rather than a silent close.

The TUI needs nothing: it renders tool_call as a result panel and carries
no pending state, so it gets the fix through the backend alone.

subagents: a worker persona, a live prompt fallback, and guards that scale ...

- assistant and designer_3d get their own subagent_system_prompt.txt; without
  one a sub-agent ran on the orchestrator's prompt, or since ce55325 on nothing
  but the scope anchor, and rarely finished the step it was given
- build the sub-agent system prompt in one place; the documented fallback to
  profile.system_prompt had become dead code
- the thinking-stall guard killed healthy cloud reasoners at 12k chars after
  13-35s; raise the char bound to 48k, keep the 60s wall, and make both settings
- a 300s run cap cut long ops work in half; raise it to 600s, configurable

sessions: the info dialog shows a pending recall ...

GET /sessions/{id}/info now carries the session's pending recall, and
GET /sessions/{id}/recall answers in one shape ({"recall": {...} | null})
instead of a flat object when found and a wrapper when not. The lookup is a
single indexed row rather than the newest 50 recalls filtered in Python.

compressor: count the context the model is sent, not the history it never sees ...

The session-info modal showed "817 288 / 131 072 (100%)" for a session whose
context held a 1.47 MB code_exec result. The estimate was honest about
session.context, but build() head/tail-truncates every oversized tool/assistant
message to context_message_token_budget before the call, so the model never
received that megabyte.

truncate_oversized() moves the per-build rule to module level (the method now
delegates, behaviour unchanged) so the compressor can price the same view it
will send: compress_and_save_session stores estimate_tokens(truncate_oversized(
new_context)) instead of estimate_tokens(new_context).

Full content stays in stored history — the user still sees it, reloads are
unaffected. Regression test keeps a 1 MB tool result in the context and asserts
the stored count equals the truncated-view estimate and is >10x smaller than
the raw one.

webclient: a dialog that answers which model is serving, and from where ...

The header only named a model when the route had fallen back, so "what is
running this session, on which host, with how much context" had no answer
in the UI. The context percentage doubles as the second way in: a session
bar is a fair place to ask what the bar measures.

GET /sessions/{id}/info reads in-process state only — the active route
entry, the health chain, the profile's effective window, run and client
counts — and the dialog polls it while open. /meta and the session head
now report the route's window instead of the global ollama_num_ctx, which
had been wrong for every cloud-backed profile.

llm: routes that name a host, a backup and a window, and a profile that reads them ...

`llm.json` replaces `OLLAMA_BACKENDS_FILE`. It has targets (host and key
environment variable) and routes, where a route is an ordered chain of
`model @ target` entries, each with its own `context_window` and options. The
`default` route now starts on `glm-5.3-flash:cloud` and backs it with
`qwen3.6:35b-128k` on the home host, both at 131072 tokens, so a failover no
longer changes the budget the compressor and `num_ctx` are accounted against.
Embeddings get their own `embedding` route instead of a pinned host.

The engine classifies a failure (network, quota, auth, server, missing model),
cools that entry and moves the call down the chain; a background probe watches
only degraded targets and lets the next turn return to the primary once it is
healthy — it never interrupts a running turn. A profile joins a route with
`llm_route`, and the loader now actually reads it: `llm_route` and
`context_window` were declared on `AgentProfile` and written into every
`config.json` while the loader skipped both, the same silent no-op that
`is_admin_only` used to be, so every profile ran the legacy model list at the
global window. The writer emits them too, or an admin edit would erase the
route from disk.

The turn's resolved entry travels to the UI: the first chunk is stamped with
model and target, `model_info` carries `fallback` when a backup answered, and
the chat header shows a chip only then. `serving_primary` compares the target
as well as the model name, because a route may run one model on two hosts.
`/health` gains an `llm` block with the active entry, target states and missing
models.

`deploy/ollama/` keeps the local model's definition honest: MTP off (slower at
every context on this host), `presence_penalty 0` (the base model's 1.5 breaks
agentic tool calls), and `num_gpu 40` — Ollama's automatic fit claims a 128k
f16 slot fits and then dies mid-prefill with CUDA OOM; 40 of 42 layers leaves
~400 MiB and decodes at ~77 tok/s, 38 layers collapses to 48, and 39 OOMs.
q8_0 KV measured 37 tok/s against 70 on f16 here, so the cache stays f16.

navi: a token count that matches the model, a type for every file, and a socket that admits it is dead ...

Compression decided everything in chars // 3, and on this workload that sits
~1.8x below reality — 101k real prompt tokens against a ~56k estimate — so the
trigger, the target, the truncate budgets and the pre-call size check all ran
in a unit half the size of the real one. TokenCalibration now fits
overhead + chars / chars_per_token per model from the prompt_eval_count that
every call already reports, and a profile declares its own context_window
(sent to Ollama as num_ctx) instead of every model being held to OLLAMA_NUM_CTX.

Superseded summaries were never demoted in the database: save() only updates
flags for rows in session.messages, a summary is is_display=False, so it lived
in session.context alone and every reload resurrected it — prod sessions
carried up to 15 live summaries. PgSessionStore.set_context_flag() writes the
flip for those rows directly.

Five extension tables disagreed about what a file is. navi/content_types.py is
the one detector — signature, then extension, then a text probe — it decides
the MIME the download route serves, and the session file list now carries the
type the viewers branch on.

The idle receive loop answers client pings and sends its own heartbeat. A
half-open socket keeps readyState at OPEN and never fires close, so traffic
coming back is the only proof the connection still works.

mcp: a run says why a server's tools were withheld, instead of silently lacking them ...

A profile can declare a BYOK-gated server the user has no personal key for. The
manager refuses such a call and build_tool_list drops the server, so the tool
list stays honest — but the system prompt kept naming the server with no
explanation at all. The model knew it existed, held none of its tools, and had
nothing to tell the user, who concluded the assistant simply could not do the
task.

The reason was never stored: it is the complement of what visible_servers kept,
so _apply_mcp_scope derives it from that same value (byok_servers() - allowed)
rather than reading the key store a second time and risking a disagreement with
the first read. It rides in current_keyless_mcp_servers, next to the scope
variable, and _mcp_context_msg names each withheld server where it lists it,
followed by the reason and a pointer at Settings -> MCP keys.

Both ContextVar tokens are returned and reset together; without that the two
existing reset sites leak or raise. The variable defaults to empty, so a run
that never set it (a peer run out of a fresh task) claims nothing.

mcp_status said "without one this server runs on the shared default", which is
false for an ordinary account — the shared credential is admin-only.

memory: a fact written under a new spelling of its key merges into the one that survives ...

The extractor invents a key every time it meets the same fact — `language`,
`preferred_language`, `ui_language` — and the upsert keys on the exact string, so
the base grew several spellings of each durable fact. On prod, 24 facts across 11
clusters share a key once noise prefixes are dropped, 10 of them under the exact
same value.

upsert_fact(dedupe=True) now compares the *tokens* of the normalized key — set
equality, not overlap, and never the value: `ui_language` and `language` are one
key while `dark` and `dark_ui` are two. The row already holding the writer's exact
key wins; otherwise the incoming one does. Losers are marked, not deleted: each
keeps its key and its value and gains `merged_into`, so the merge is reversible
with one UPDATE. Off by default, so the memory tool still writes exactly the row
it asked for.

Every read path filters `merged_into IS NULL`, and the mark is reachable only
where it is meant to be — get_all_facts/fact_count gained include_merged, and
GET /admin/memory?include_merged=true exposes it. search_facts deliberately has
no such flag: the injection path must not be able to see a hidden row. Forgetting
a key drops its hidden aliases first, so none is left pointing at nothing.

Three fixes that fell out of the same read:

- expires_at was overwritten with EXCLUDED.expires_at on every upsert, and the
  extractor always passes None — so it silently cleared any expiry the agent had
  set. Both DO UPDATE clauses now COALESCE.
- The ILIKE fallback was the one read path that served expired facts.
- The `list` action promised "all stored facts" and returned categories. The code
  was right and the description was wrong; the description now says so.

The extraction prompt stopped allowing infrastructure — servers, hosts, addresses,
running services — while the agent and the planner already forbade it; that
loophole is where 423 technical facts came from. It also forbids secrets outright,
and asks for stable snake_case keys. Expiry is now documented as being for facts
with a known end, not for facts you merely doubt.

Facts reach the prompt with their age ("updated 34d ago") alongside confidence:
nothing expires on a clock, so the injection says how old a fact is and lets the
agent weigh it.

mcp: every wait on the MCP path gets a bound, and a failed reload keeps what it had ...

reload_tools kept dying on prod, and only a restart cured it. Nothing on the
MCP path was bounded: ClientSession was built without read_timeout_seconds, so
the SDK waited forever (anyio.fail_after(None)); _send awaited a future with no
timeout; disconnect() awaited the shielded runner with no timeout; and
disconnect_all() walks the pool one client at a time, so a single wedged client
blocked every reload that came after it. A cancelled caller only cancelled
itself — the runner stayed wedged, which is why it was every reload, not one.

Bounds, in two deliberate layers. The SDK read timeout ends a request nobody
answers *inside* the runner, so the client stays usable; the caller-side bound
is the backstop for an op queued behind a stuck one, and it aborts the runner
rather than leaving it for the next command to queue behind. connect 60s,
list_tools 45s, call_tool 300s (SDK 240s), session read 30s, disconnect 15s,
each overridable, each logging the culprit.

disconnect() and disconnect_all() no longer swallow CancelledError — that is
exactly how the reload budget stopped being enforceable.

The registry is now rebuilt in one step: collect every server's tools first,
then commit the whole set in a single synchronous swap. clear_external_tools()
ran *before* the rebuild, so a rebuild that failed halfway left the model
holding MCP tool names that resolved to nothing. A failed pass now leaves the
previous registrations untouched, and the reload phase carries its own budget
so the tool always returns a report.

Two more defects fell out of the audit. The streamable_http transport was
handed a bare httpx.AsyncClient, whose 5s default read killed the GET stream of
an idle server — six of the nine prod servers are streamable_http. And
register_mcp_tools called log.warning() in a module where log was never
defined, a NameError on precisely the failure path it was meant to report.

sessions: a pasted image lands in the session directory, and the notice says where ...

An image pasted into the chat lived only in the transcript. The system prompt of
every session names `session_files/<id>/` as the place its files live, so an agent
looking for "the attached images" went there, found nothing, and spent turns on
fruitless `find` calls before giving up.

Now the agent's non-hidden turn writes each pasted image into that directory as
`image_N.<ext>` — the extension sniffed from the bytes, the bytes stored as they
arrived, so a PNG keeps its alpha channel. Numbering runs across the session, and
a picture sent twice reuses the file already holding it.

The paths go into the hidden context copy of the message: images past the context
budget are not sent to the model at all, so the notice names them for `image_view`
instead. The line is written by the agent rather than the websocket handler, which
is the layer that knows the names the files actually took — including for a message
queued behind a running turn.

A save that fails is logged and skipped: it never costs the turn.