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.

synapse: route an event by the role of the account it was delivered for ...

The dispatcher's choice of profile was checked against a hardcoded "user" role,
so an admin's own events could never reach the admin-only profiles their routing
instruction named. The event fell to whichever profile all() happened to yield
first — a profile without the tools the event needed — and the instruction looked
like it did not work. The role now comes from navi_users for the account the event
was delivered for (_user_role, failing closed to "user" and logging), and it is the
role the run carries as current_user_role.

Alongside, in the same path:

- the dispatcher is offered each profile as "id (name) — description", so a rule
  that names a profile in words has something to match it against;
- an unroutable event lands on a role-aware fallback (secretary for an admin,
  assistant otherwise) instead of the first of all(), and every substitution is
  logged as synapse.reaction_profile_downgraded with both ids;
- session_key is folded to one spelling — whitespace collapsed, lowercased — so a
  chat cannot fork on "TG:42" against "tg:42";
- the conversation rule is read from whichever instruction document states it,
  not from the dispatcher one alone.

A refused profile is still refused: hidden stays hidden for everyone, and an
admin-only profile stays out of reach for a regular user's event.

synapse: a reaction continues its conversation, and the request becomes a record ...

Every Synapse event used to open its own service session, so two messages of
one Telegram chat landed in two unrelated chats. The dispatcher now returns a
`session_key` — derived from the payload according to the user's routing
document — and a later event carrying the same key joins the session that
opened it, within `reaction_session_ttl_minutes` of idle time (0 restores the
old always-fresh behaviour). A session that is still running is never joined:
`create_run` overwrites `state.run`, so a second run would orphan the first
one's subscribers — a busy thread gets a fresh session instead. When the
dispatcher names a different profile for the event, the session's profile is
switched through `set_profile`, never through `save()`, which does not rewrite
`name` or `profile_id`: the thread keeps its name and its history.

What the user reads in the transcript is now only the request. The event
record is a system message (`metadata.source == "synapse_event"`) drawn by a
new SynapseEventNotice.vue in the visual language of the recall badge, and the
framing plus the JSON envelope become a hidden user turn written once, on the
first event of a thread — not on every one. The instructions from Settings
reach the model through a new `current_reaction_instructions` ContextVar and
the `[Reaction session]` block of the system prompt, so they appear in no
transcript at all. The stored system record still has to reach the model, so
ContextBuilder's allowlist becomes `_LLM_SYSTEM_SOURCES = {task_note,
synapse_event}`.

The dispatcher gets a document of its own — how to route, rather than what to
do — stored beside the reaction instructions, versioned separately, and edited
in its own field in the settings panel. `synapse_instruction_versions` gains a
`doc` discriminator, so existing rows keep `'reaction'` and their history
verbatim. The new columns are added by `_MIGRATE`, and the index over `doc`
lives there too: `_DDL` runs first, so an index over a column the migration has
not added yet kills the whole batch on an existing database — which is what
the prod database is. A guard test keeps it that way.

Tests: 1776 passed, 1 skipped (backend), 242 passed (webclient, 28 files).

profiles: restricted profiles for ordinary users, and a role gate that holds ...

`is_admin_only` was checked in one place out of nine and was not read from
config.json at all, so all seven profiles were reachable by every account with
role `user`. The flag now lives in config.json — the file is the baseline, a
`profile_overrides` row still wins on top of it — and one predicate,
`admin_only_blocked`, is the single place the rule is expressed. The nine
surfaces that list, switch to, spawn or resolve a profile all consult it:
`POST /sessions`, the WebSocket, switch_profile, list_profiles, the system
prompt's "Available profiles" block, spawn_agent and the Synapse reaction
runner. The prompt cache is now keyed by (profile, role), so a user's prompt
can never be served an admin's profile list.

The seven existing profiles (developer, discuss, dispatcher, modeler_3d,
navi_code, secretary, server_admin) are marked admin-only. Three new ones take
their place for ordinary users: assistant, designer_3d and coder. They share one
native tool set — ssh_exec, peer, reload_tools, create_mcp_server, test_mcp_tool,
image_view and gmail are withheld — and differ only in system prompt, model and
MCP groups. navi-web's raw `request` group, and the whole of gnexus-creds and
tgclient, are withheld too.

MCP per-user keys gain the missing half of the rule: a server that declares a
`user_key` slot is refused to anyone but an admin who has no personal key, and
is left out of their tool list entirely, instead of quietly falling back to the
owner's credential and appearing as a tool that cannot work. The refusal names
the server and points at Settings.

Also closes `GET /agents/prompts`, which served every profile's system prompt to
anyone, with no user dependency at all.

The accepted residual risk is written down in docs/profiles.md: the working
directory is a convention, not a sandbox.

subagents, secrets: an honest result contract, and two leaks closed ...

A pass over the sub-agent system, plus the two findings from the previous one.

Leaks:

- `redact_args` / `is_sensitive_key` (navi/tools/_internal/redact.py) mask
  credential-shaped keys at every log site that prints tool arguments — by key
  name, for all tools, so a future MCP tool is covered without declaring
  anything. ssh_exec passwords were sitting in the prod journal in plaintext.
- `mcp_servers.d/*.json` are tracked in git and carried the scaffolding
  machine's absolute paths, so navi-3d/navi-web silently vanished from the
  toolset off that machine. They now hold project-relative paths, resolved in
  `McpClient.open_transport` — not at load time, because create_mcp_server
  round-trips every config file through save_mcp_servers.

Sub-agent result contract:

- `run_ephemeral` returns a `SubAgentOutcome` (text + status: ok, timeout,
  max_iterations, thinking_stall, user_stop, context_overflow) instead of
  `(text, bool)`. Every exit routes through `SubAgentRunner._finish`, so token
  accounting is never dropped and a partial run always carries a progress
  report — the stop-after-stream path used to return a bare string.
- `spawn_agent` renders its header from the status. Before this, every short
  run reached the parent as "hit iteration limit", and the parent repeats that
  diagnosis to the user — a timeout was reported as an iteration limit.
- `ContextTooLargeError` no longer kills a sub-agent: the parent compresses its
  own context, a sub-agent has no session to compress into, so it stops with
  `context_overflow` and hands back what it did rather than surfacing as
  "Sub-agent failed: …" with the whole run discarded.
- The result is capped at 8000 chars (head and tail kept) so a long run cannot
  flood the parent's context, and tool arguments in the progress report go
  through `redact_args`.
- The navi-3d calling convention left the runner for the server's own tool
  schemas — a generic runner should not know one MCP server's path shapes.
- Validation moved inside spawn_agent's try: a bad argument is a clean
  ToolResult the model can read, not an exception in the parent's loop.

Tests: 1636 passed, 1 skipped. New: SubAgentOutcome paths and headers, result
truncation, context overflow, clean argument failures.

tools, mcp: a description is a hook, not a manual ...

With 24 tools in a profile the schemas are the largest fixed block of every
request, and they are paid for whether or not the tool is called. Eleven of
them carried a manual's worth of prose in their `description` — examples,
error tables, "common mistakes", the mechanics of the file areas. Reading
them against the manuals first, the manuals already held all of it, and
richer: this was duplication sitting in the one place the model cannot avoid
reading, not knowledge that had nowhere else to live. So the descriptions
became selection hooks — what the tool is for, when to pick it over its
neighbour, and the one rule that cannot wait — and the detail moved to
tool_manual("<tool>"), which costs nothing until the tool is actually in
play.

filesystem, spawn_agent, schedule_recall, manage_recall, scratchpad, plan,
share_file, content_publish, reflect, todo and peer. Descriptions drop from
18.4 KB to 11.5 KB, those tools' schemas from 25.5 KB to 16.8 KB, and a
profile's whole native toolset from 38.3 KB to 29.6 KB (server_admin,
developer; secretary 33.7 → 25.0, navi_code 32.9 → 25.9, tool_developer
37.9 → 29.2, modeler_3d 33.4 → 24.8). Stated as a budget rather than a byte
count: roughly 4k tokens of the model's window per request, back.

Two things were deliberately not shortened. The filesystem edit ladder
(`edit` → `edit_lines` → `smart_edit`, last resort, costs an LLM call) and
todo's mandatory `validation` on `done` both exist to prevent an extra
round trip; a shorter description there would trade tokens for turns. What
todo lost is the prose around the rule, not the rule. `enabled.json` was
left alone too: weather, gmail and get_current_datetime are opt-in and the
DB shows 12, 16 and 40 calls, so they are opt-in tools that are actually
used.

peer had no manual at all, which is how this surfaced: its description was
the only documentation the tool had. manuals/peer.md is now written from
the tool and its /peer route — the ask/status/list actions, the fact that
the hive is only a phone book and the question travels peer-to-peer, the
one-concurrent-answer semaphore, the two independent recursion guards, and
the six error codes.

The MCP server instructions leave the system prompt the same way. They were
8 KB of always-present prose per request, most of it workflow detail the
model only needs when it is about to use that server. Each server now
contributes one line — a new optional `summary` field in mcp_servers.d/*.json,
defaulting to the first sentence of its `instructions` — and
tool_manual("<server>") returns the full text plus the tool names the server
declares. server_admin's MCP block goes from 8.3 KB to 1.3 KB, secretary's
from 8.1 to 1.2. tool_manual learned to answer for a server (a tool name
still wins over a server of the same name, and dash/underscore spelling is
tolerated), and the index names the servers alongside the tools.

One judgement call worth recording: gnexus-book's and navi-web's
instructions end in an absolute "NEVER bypass these tools with filesystem,
terminal or code_exec" rule. Moving that on demand would have been a
behavioural regression dressed as a token saving, so those two `summary`
fields carry the prohibition verbatim alongside the hook, and it stays
always visible.

docs/tools.md gains a section on the split (a description is paid for every
request, a manual only when the tool is used), which is where the next
person will look before trimming a description back into a manual.

Full suite green (1577 passed, 1 skipped). No frontend change and no new
dependency, so the deploy is a pull and a restart.

agent: a profile switch now survives the turn that made it ...

switch_profile repoints the session with one narrow UPDATE, and the run's own
save() — which lands right after every tool batch, the switch's own batch
included — wrote its stale in-memory profile back over it. The post-turn
reload then read the row it had just clobbered, saw no change, and left the
turn bound to the old profile's tools: the agent called exactly the tools
switch_profile had just advertised ("Newly available: ssh_exec, terminal, …")
and got "tool not found" for every one of them. Prod shows both halves — the
session that reported two switches kept the profile it started with, and
agent.profile_reloaded appears zero times in the whole boot.

The reload now compares against the profile the run is bound to
(bound_profile_id), which set_profile cannot move, instead of against
session.profile_id, which the store either rewrites or mutates to the new
value — making the old guard false in both branches by construction. And
save() leaves the profile_id column alone on conflict: set_profile owns it
after creation, the same way name/created_at are owned elsewhere.

A tool the live tool_map does not hold is now logged (tool.not_found, with
profile and live tool count). It previously left no trace at all in the
journal, which is why this took a database dig to find.

Three tests, each verified to fail without its fix: the next LLM call is
offered the new profile's tools, save() does not rewrite profile_id, and a
missing tool is logged.

reload: pick up the new code, and drop MCP tools that are gone ...

reload_tools reported success while three separate things kept it from
doing what it says.

The bytecode cache is keyed on (mtime in whole seconds, file size), so a
tool edited to the same length inside the same second as its previous
load re-ran the OLD code — the reload was real, the new version was not
live. The loader now compiles the source itself instead of consulting
__pycache__ (importlib.invalidate_caches() does not help here).

MCP registrations only ever grew: register_mcp_tools called
register_external, and unregister_external was used in one place, so a
server removed from the config or a tool a server stopped exposing
stayed in the registry and failed only when the model called it. Reload
now clears external tools and rebuilds them.

enabled.json naming a tool that failed to load (the gmail/html2text case)
was visible only as a log line. It is now part of the report.

The reload itself moves to navi/core/reload.py, one implementation shared
by the tool and the admin route, so the two cannot leave different
toolsets behind. list_tools.py also read enabled.json through its own
cwd-relative path, which made it disagree with the real toolset, and the
class-based loader rejected execute(self, params, ctx=None) — the shape
every built-in uses.

switch_profile: run the switch alone, then the batch on the new tools ...

A tool called in the same batch as switch_profile was still dispatched against
the old tool_map and died with "tool 'X' not found" — reload_tools did, twice
in one session. The batch is now split: the switch runs first, its
ProfileSwitched event is watched for the target profile, the tools are
re-resolved from it, and the remaining calls run against the new set.

The turn's own profile binding is deliberately left alone — the
end-of-iteration reload in run_stream() is what rebinds profile/llm/schemas,
and pre-setting session.profile_id here would make it skip that step.

Allocate message sequence numbers in the DB, not in memory ...

save() numbered new messages from Session.db_next_sequence, a counter read
when the session was loaded. Any second writer of the same session handed
out the numbers it still believed were free, and the two inserts collided on
UNIQUE(session_id, sequence_number) — in production, when switch_profile
loaded the session mid-run and saved it back. The turn died with
"Internal error: duplicate key value violates unique constraint".

The range is now claimed inside save()'s transaction with a single
UPDATE ... RETURNING, so concurrent writers serialize on the session row,
and GREATEST() seeds the pre-counter sessions whose next_sequence is still 0
instead of relying on a racy max()+1 fallback in memory.

switch_profile itself no longer saves a session at all: it repoints the row
through a narrow set_profile() UPDATE, which keeps it out of the running
turn's way. It runs mid-turn on a session the turn still holds, so saving a
second copy from there was the collision in the first place.

Await the session pool in notify and three neighbours ...

PgSessionStore._get_pool() is async, and four call sites passed its coroutine
straight into a store constructor, so the first query inside died with
"'coroutine' object has no attribute 'fetchrow'": notify always failed, the
reaction runner never got past reading its settings, synapse_instructions was
unusable, and the BYOK resolver caught the AttributeError and silently fell
back to the default credential.

The tests missed it because each one mocked the store or the pool provider
away; the new ones run on a fake asyncpg pool with nothing faked below the tool.

synapse reactions W3: reaction runner, notify tool, gateway hookup ...

- navi/synapse/reactions.py: fire-and-forget reaction run — per-user
  reactions_enabled gate, one-shot dispatcher meta-pass (visible profiles
  only, hidden dispatcher, secretary fallback), special=True session
  headed by the resolved profile, headless agent run via the orchestrator
  run registry, finalise: app push per completion_notify +
  low-level reaction.finished/failed event to Synapse
- navi/synapse/outbound.py: source-side emit (syn_* key from env,
  silent when not configured)
- navi/tools/notify.py: notify tool — message + level (info/warning/
  intervention), both legs per push_target, honest delivered/skipped result
- push/service.py: notify_custom without cooldown
- gateway: verified + non-duplicate delivery schedules the reaction
- profiles: notify added to native tools (7 profiles)

synapse reactions W1: special sessions flag + per-user reaction settings and instruction history

navi: session naming folds agent's final replies into context

compressor: single-summary invariant — old summaries always fold into the fresh one ...

partition_messages now strips is_summary messages out of the turn structure
before grouping: a summary is already-compressed history, not a user request.
Previously its role=user made it a turn head, so midturn compression kept the
old summary verbatim beside the fresh one (head kept + new summary added),
and adaptive swap resurrected summary turns whose text pattern-matches the
importance heuristics. This piled stale summary copies into the context
(one session accumulated 6 copies = 120k of 239k chars) and made compression
no-op for whole stretches before a lucky pass finally shrank it.

Also folds summaries on the early-exit path so two stale summaries
consolidate via the meta-summary, and fixes the demotion loop in
compress_and_save_session to cover session.context too: after a reload,
context-only rows (is_display=False summaries) were absent from
session.messages and never demoted, so is_context=True persisted in the DB
and the copies survived into the loaded context.

background tasks: review fixes B1-B11 batch ...

- stop mid-batch cancels in-flight tools (B1) and keeps real results
  of already-finished ones, mixing them with synthetic stopped notes
  in call order (B2)
- queued messages: headless drain publishes session_sync (B4), the
  run's teardown broadcasts session_sync to other sockets but not the
  owner socket (prevents double reload) (B5)
- task_update notifications are chained per task so late running
  updates can't overtake the terminal one (B7; client drops late
  re-flicker of a terminal task chip) (B8)
- client: ui_component results reference the owning message via
  card.parentMsg instead of a stale msg reference (B6)
- tasks cancel reports the real outcome after a bounded wait instead
  of an optimistic 'cancelled' (B9)
- session delete cancels its running background jobs and drops
  pending result notes (B10)
- subagent tool loop routes background:true calls through
  ToolExecutor._maybe_background like the main loop (B11)
- tests for all of the above; docs/tasks.md stop/cancel semantics

local mode isolation: anonymous user is a NULL-owner user, not an alias ...

Local mode (NAVI_AUTH_ENABLED=false) used an anonymous admin whose admin
role widened every session listing to ALL users' chats — acceptable only
while assuming a private database, but a real leak on any shared one.

- session listings (list_all/list_page/count_all/search_list): new
  scoping rule — non-admin with user_id=None sees only user_id IS NULL
  rows; named owner unchanged; admin unchanged
- session create/list routes: owner id is None in local mode (sessions
  are persisted ownerless), admin listing flag resolved only when auth
  is enabled
- check_session_access: local mode allows only NULL-owner sessions —
  a user-owned session by id is 403
- debug/admin endpoints (require_admin) stay reachable in local mode
- prod DB: 30 stray user_id='anonymous' rows reassigned to NULL owner

Tests: store scoping unit tests, local-mode integration tests (sidebar
filtering + access by id), updated check_session_access unit tests.
Full pytest 1239 passed, 1 skipped.

chat history pagination: paged load instead of full-session fetch ...

Backend:
- PgSessionStore.get_meta — light session head (sessions row + COUNT),
  no message payload
- PgSessionStore.get_history_page — page of display history (archive ∪
  hot UNION), newest-first page, oldest-first result, limit+1 peek for
  has_more; each message stamped metadata.display_index (ROW_NUMBER-1
  over the full display history) so client ids stay stable across
  paged loads; dangling-tool-call repair keeps real ranks by
  sequence-number lookup
- GET /sessions/{id}/meta and GET /sessions/{id}/messages/history
  (before_seq cursor, limit ≤200)

Webclient:
- loadSession/reloadSession fetch meta + newest 200-item page
  (Promise.all), archive state from the page itself
- buildMessageList ids (h_*) and rawIndices use the global
  display_index — feedback keys and search-jump stay stable when
  history is paged
- scroll-up loads older pages through the same history endpoint
  (archive table alone misses sessions whose threshold never moved)
- search-jump pulls pages until the target index is covered (bounded)

Tests: store unit tests (ranks, has_more peek, placeholder shift) +
vitest updates; full pytest 1231 passed, vitest 94 passed.

context: task-note messages survive build()'s system filter ...

E2E caught a real delivery bug: the completion note was drained and
persisted (agent.task_notes_drained logged) but build() filtered out all
system-role history from session.context, so the LLM never saw it — the
agent could only get detached results via tasks check. Notes now carry
metadata source=task_note (kept by build) and is_display=False (clients
already show live task_update events). Verified live: agent reads the
note verbatim without calling any tool.

e2e fixes: tasks tool in profiles, bg timeout lift, stepIcon fix, queue semantics docs ...

E2E findings addressed:
- tasks tool was registered but not in any profile's tools.agent.native —
  added to all six profiles (agent could not check/wait/cancel bg tasks)
- detached terminal/code_exec/ssh_exec runs without explicit timeout are
  lifted to 300s: foreground defaults (20/30/60s) marked long commands
  'completed' with partial output while the process still ran
- ToolCard.vue: define stepIcon(status) — template referenced it but the
  function was missing (render crash on task_update step)
- message_queued reachability documented: WS read loop is sequential, the
  queue path is only reachable from a second socket/headless recall

agent parallelism: background tools, bg spawn_agent, parallel tool batches, message queue ...

- backgroundable tools (terminal/ssh_exec/peer/spawn_agent/code_exec) detach
  via TaskManager with per-session/global/spawn caps, rate limit and TTL
- completion delivery both ways: task_update event (out-of-band) + pending
  notes injected into the next turn; tasks tool (list/check/wait/cancel)
- parallel tool-call batches (profile-level gate, off by default) with
  ToolStarted up-front, tagged event mux, call-order results, one save,
  plus dangling tool_call repair on session load
- user message queue instead of busy error: message_queued frame,
  back-to-back drain on the same socket, headless fallback on disconnect
- pin mcp<2 (v2 renames FastMCP with breaking API changes)
- docs: tasks.md (new), websocket/api/agent/config updates, tasks manual,
  spawn_agent background param, persona contract section

PWA: installable webclient, offline shell, web push ...

Installability:
- public/manifest.webmanifest (standalone, theme #16161E) + PNG icons
  generated from logo.svg (regular + maskable, served via /images mount)
- index.html: manifest link, theme-color, apple-touch-icon

Offline shell:
- hand-rolled sw.js (no workbox): navigation = network-first (3s race)
  with cached-shell fallback + background refresh (a stale cached shell
  would 404 on entry chunks after a deploy); /assets/* cache-first
  (content-hashed); /images/* cache-first capped; /api,/ws,/auth,/push,
  /content pass-through
- vite closeBundle plugin stamps __NAVI_BUILD_VERSION__ (digest of
  index.html + asset names) into dist/sw.js; sw.js served no-store so
  every deploy reactivates the SW and activation evicts old caches
- SW registration in main.js, PROD only (dev HMR untouched)
- OfflineBanner (useOnline composable) over the app shell

Web push (VAPID, pywebpush):
- navi/push/ package: push_subscriptions table (postgres, boot-time DDL),
  PushSubscriptionStore, PushService (async fan-out, to_thread sends,
  404/410 prunes dead endpoints, per-session cooldown)
- routes: GET /push/vapid-key, POST/DELETE /push/subscribe (auth-gated)
- trigger in orchestrator run_agent + run_recall: push on StreamEnd when
  no WebSocket client watches the session; fire-and-forget, never
  disturbs the run; anonymous fallback only when auth is off
- client: usePush composable + Notifications settings panel (enable/
  disable via PushManager.subscribe with the server VAPID key)
- notification click focuses the app at /#<session_id> (hash routing
  opens the right chat); payload body is a markdown-stripped <=140-char
  preview

NAVIVAPID keys empty = push fully disabled (graceful, like other optional
integrations). dist/ artifacts committed per repo convention.

Tests: pytest push store/service/routes/trigger (+23), vitest usePush
(83 webclient tests green). Full suite 1143 passed.

planning: fix LLM-output failure modes that killed every plan ...

Production planning_logs showed three ways a plan silently dies:

1. glm-5.3-flash wraps the whole answer in a structured envelope
   (response:unknown{value:...}<tool_call|>) — the planner only knew the
   gemma4 thought<channel|> artifact, so the wrapped analysis/plan failed
   to parse. The stripper now unwraps response:<type>{value:...} envelopes,
   including a truncated (unbalanced) one.
2. Empty content with spent completion tokens: glm on ollama-cloud puts
   the whole answer in the thinking channel on some calls. Both planning
   phases now fall back to thinking when content is empty; the debug log
   records which channel served (source: content|thinking).
3. phase3_timeout at the shared 120s LLM_COMPLETE_TIMEOUT: planning now
   has its own PLANNING_LLM_TIMEOUT_SEC (default 240).

Unit tests cover the envelope unwrap (balanced/truncated/inner braces),
the thinking fallback in both phases, and the clean-fail path.

compress: route every trigger path through compress_and_save_session ...

- CompressionWorker keeps the real-token gate (ctx.context_tokens) but now
  delegates to compress_and_save_session(reason="postturn") — it used to
  call compress_context directly with its own save logic and silently
  no-op'ed on any summarizer failure, leaving the session over the
  threshold until the next turn's gates
- pre-turn and /compact gates use real_baseline_estimate instead of the
  chars//3 heuristic (mid-turn already did), so code-heavy contexts
  compress in time instead of tripping ContextTooLargeError
- summarizer input keeps head + tail instead of head-only truncation,
  which silently dropped the newest summarized messages closest to the
  keep window
- critical tool results over the 4000-char budget keep head+tail halves
  instead of collapsing to the 800-char non-critical preview (a 3999-char
  read survived whole, a 4001-char one lost 98%)
- images count 500 tokens each, not 500 per message carrying them
- _plan_compression is the single partition decision shared by
  compress_context and would_compress, so the dry-run prediction cannot
  drift from the real path; the trigger reason is now logged
  (preturn/midturn/postturn/forced)

planning: replace the mandatory pre-turn gate with an agent-invoked plan tool ...

- Phase 2 (critique) retired; observe/MODE and the top-level DIRECT shortcut
  removed (sub-agents keep DIRECT for trivial subtasks)
- replan merged into plan: `reason` (+ updated_goal) packs the re-plan
  context, without a reason it is fresh planning
- both phases call the LLM with think=False — cloud reasoning models leak
  their chain-of-thought into structured output on non-streaming calls —
  and strip the gemma "thought<channel|>" content artifact
- Phase 1 conversation windowed to ~20k chars, newest-first with the
  original task pinned (28.7k-token prompts came back as 1-token output)
- confirmation by COMPLEXITY: the tool result tells the agent to present
  complex plans and wait; sub-agents keep the execute-now injection
- PlanningStatus/PlanReady reach the UI mid-turn via current_event_sink
  (ctx.event_sink is None in the agent loop — events silently never
  reached the WS before)
- agent gate, casual-message detector and adaptive re-plan nudges removed;
  anti_stall keeps only stall detection

compress: raise budget 6000 + richer prompt detail + input headroom ...

Increase the compression summary budget and push the prompt toward more
detail and completeness, so the agent keeps more durable facts when old turns
are folded into a summary.

Budget (output):
- context_summary_max_tokens 4000 → 6000 (global default).
- navi_code compression_max_tokens 4000 → 6000 (profile override).

Prompt (base, all profiles):
- New 'Intermediate Findings' section: durable facts from read/grep/log/terminal
  that informed decisions (was forbidden by 'do not preserve intermediate
  reasoning'). Existing sections get richer instructions (Active Files: key
  symbol affected; Decisions: rationale + trigger; Completed: step + file +
  verification outcome; Errors: snippet + what was tried + outcome).
- Loosen 'tight bullet points' → 'specific and complete; prefer listing over
  generalizing — losing a fact costs more than a longer summary'.
- 'Do not write implementation code' → 'preserve short critical code verbatim
  (final signatures, one-line fixes, key config lines); drop long blocks'.

Input headroom (so the summarizer can SEE what it must preserve):
- Non-critical tool result preview 300 → 800 chars.
- _MAX_SUMMARY_INPUT_CHARS 24000 → 32000 (fits 65k–128k windows with 6k output).

Profile prompt (navi_code/compression_prompt.txt): preserve short diagnostic
snippets verbatim (exact error line, failing assertion, one-line fix) instead
of blanket 'do not preserve long output'; keep code signatures + short key
snippets, not just names.

Tests: Intermediate Findings section in base prompt; 'short critical code'
phrasing; non-critical preview keeps 800 chars; meta-summary threshold raised
(2×6000 chars). Fixed flaky test_attach_session_seeds_context_fill (relied on
_startup worker timing after Etap 4 lengthened attach; now explicit attach
like its sibling + mock get_todos to skip a 30s×2 real-network timeout).
Full suite: 1008 passed, 1 skipped.

compressor: target hysteresis — shrink to 65% after compression, not just below the trigger ...

Compression triggered at 90% but left the context just under the trigger, so a
fixed keep_recent (navi_code: 12 turns ≈ 104 messages) re-triggered a couple
of messages later — the context yo-yoed at the trigger line instead of gaining
headroom. Add context_compression_target (0.65): in turn-based (preturn) mode
compress_context shrinks keep_recent until the verbatim kept region fits 65% of
the window, folding the extra turns into the same single summary LLM call; a
safety net in compress_session token-budget-truncates if the kept region alone
still exceeds the target (midturn kept a huge in-flight turn, or the
keep_recent floor can't fit). Trigger 90% → target 65% leaves real headroom.

Автономность мелких моделей: OUTPUT DISCIPLINE, milestone-todo, перехват финала хода ...

Два последовательных захода над одной проблемой — мелкие модели (12–30B) в navi_code
теряют автономность на трудных шагах: «остановился поболтать» вместо действия и
слишком большое расстояние между пунктами плана.

Заход 1 (З1–З3):
- З1: OUTPUT DISCIPLINE в системном промпте navi_code — «act, don't announce»,
  без few-shot антипримеров. Контракт хода: ответ без tool_calls = конец хода, поэтому
  объявление намерения текстом убивает автономность; правила заставляют вызывать
  инструмент в том же ходе.
- З2: плоский todo + метка группы milestone + декомпозиция. _parse_plan_steps
  возвращает list[tuple[milestone, text]]; milestone — метка группировки (не сущность,
  без статуса), «done» вычисляется при рендеринге; подшаги = больше плоских шагов
  (без вложенности). TUI side-panel группирует по milestone (плоский фолбэк при пустом
  milestone). Plan depth: max 15→20 + правило декомпозиции.
- З3: adaptive re-plan «длинный шаг» — nudge «разбей шаг» при in_progress ≥ порога
  итераций без смены todo (порог 4, раньше общего anti-stall warning на 8).

Заход 2 (шаги 1–3, после cloud-теста 31b vs 12b):
Корневая структурная причина: весь спасательный механизм (anti-stall warning с явным
предложением reflect, adaptive re-plan) живёт только внутри tool-цикла — nudge
инжектируется в pre_turn *следующей* итерации, которой при «остановился поболтать»
нет (ход закрылся по return до post_turn). 31b застревала через «продолжаю
tool-итерации» → дожала до warning → спаслась; 12b — через «замолчала текстом» → мимо
всех nudge.

- Шаг 1: перехват финала хода. Если модель выдала bare-text, но в todo есть открытые
  шаги (pending/in_progress) и лимит не исчерпан — НЕ эмитить StreamEnd, а сохранить
  ассистентский текст в session.context, поставить системный nudge и continue
  (без StreamEnd, без workers — консистентно с multi-iteration tool-турами). Счётчик
  final_interceptions на AgentTurnContext, лимит final_intercept_limit (default 2),
  эскалация жёсткости (мягкий → «second stop»). has_open_steps в todo.py: пустой
  todo → False (защита casual-сообщений), failed/skipped терминальны. Профильные
  флаги final_intercept_enabled/limit в base.py + loader + admin.
- Шаг 2: жёсткий reflect-триггер в промпте — «~3 tool attempts on the same step
  without progress → call reflect IN THIS TURN (tool call, not reasoning aloud)».
- Шаг 3: открыть replan для застревания — «call replan when reflect showed the whole
  approach is dead (not one failed step, but the approach itself)».

Тесты: 874 passed, 1 skipped. Новые — has_open_steps (5), final intercept (5),
milestone-группировка, adaptive long-step nudge, парсер шагов с milestone-маркером.

cloud: navi_code model → gemma4:31b-cloud для тестирования догадки (31b признала
застревание, 12b — нет); .env cloud-host уже gitignored.

Co-Authored-By: Claude <noreply@anthropic.com>