Avoid applying text-only persist_user_message overrides to multimodal current-turn user messages. Early crash-resilience persistence mutates the same messages list later used for the API call, so clobbering list content drops ACP image blocks before model dispatch.\n\nAdd regression coverage for both text override behavior and multimodal preservation.\n\nCloses #44242
* feat(mcp): raise default tool-call timeout 120s -> 300s
Port from openai/codex#28234. Long-running MCP tools (web fetches,
sandboxed builds, deep-research servers) routinely exceed 120s, causing
spurious timeout failures. Codex bumped its default MCP tool timeout from
120 to 300 for the same reason.
- _DEFAULT_TOOL_TIMEOUT 120 -> 300 in tools/mcp_tool.py (per-server
'timeout' config override unchanged)
- update test_default_timeout assertion
- document the default in mcp-config-reference.md
* refactor: remove agent-callable send_message tool
The agent should not decide on its own to fire off cross-platform
messages or reactions. Outbound platform messaging is handled outside
the agent loop — cron delivery, the gateway kanban notifier
(dashboard-toggled), and the `hermes send` CLI.
Removes the model-tool registration only; the send engine in
send_message_tool.py (_send_to_platform, _send_via_adapter,
_parse_target_ref, per-platform _send_* helpers) is kept intact for
those non-agent callers. Drops the now-empty 'messaging' toolset and
its `hermes tools` toggle. Yuanbao DM guidance now points at the
native yb_send_dm tool.
A multi-MB message (logged bundle, huge tool dump) froze the renderer
before any paint: Streamdown runs `preprocess` + `marked` lex over the
whole string synchronously in a useMemo, an uninterruptible long task
that no try/catch or content-visibility can help (our JS runs before the
browser ever skips layout). Tiered fix:
- Message gate: past 200KB, bypass markdown entirely and render the raw
text in `content-visibility:auto` line-chunks — synchronous work is
bounded to a string split, the browser virtualizes layout natively,
and every line stays in the DOM (selectable, find-in-page).
- Code-block budget: past 3k lines / 150KB, skip Shiki (which emits a
span per token) and render plain, chunked the same way.
- Collapse/expand: a reusable ExpandableBlock clamps code blocks and the
huge-text fallback to a 120px preview with a gradient + chevron,
expanding to 300px. The inner element is always a scroll container so
the content-visibility chunks stay lazily laid out in both states.
No content is ever dropped; the copy button (card header) always yields
the full block.
restore_skill() falls back to p.name.startswith(f"{skill_name}-") when no
archive directory matches the requested name exactly. That fallback is meant
to catch the timestamped duplicate archive_skill() writes on a name collision
(<skill>-YYYYMMDDHHMMSS), but the bare prefix also matches any unrelated
archived skill named <name>-something. So restoring "git" can pull an archived
"git-helpers" out of .archive/, rename it to "git", and report success: the
requested skill is not restored and the sibling is gone from the archive.
Constrain the fallback to the exact suffix archive_skill() produces, a 14 digit
timestamp. The exact-name match and the recursive nested-archive walk are
unchanged, so nested and timestamped restores still work; unrelated siblings no
longer match.
Fixes#47647
The OpenAI device-code login (POST auth.openai.com/.../deviceauth/usercode)
had no retry or 429 handling — a transient throttle from OpenAI surfaced as
a bare "Device code request returned status 429" with no guidance, reading
as a hard login failure.
- Retry the device-code request with capped exponential backoff (honoring
Retry-After), up to 4 attempts.
- On persistent 429, raise a clear AuthError tagged CODEX_RATE_LIMITED_CODE
(classified transient, not a credential problem) with a wait hint.
- Apply the same 429 classification to the token-exchange step (same bug
class).
Unrelated to PR #47399 (Responses-API cache headers); this is the OAuth
device-code path in hermes_cli/auth.py.
get_runtime_status_running_pid() validates liveness with a local
os.kill(pid, 0) probe. In /api/status the runtime record can be the
REMOTE health-probe body (cross-container), whose PID belongs to another
host and is display-only — probing it locally is wrong and trips the
test live-system guard (os.kill on a PID outside the test subtree).
Run the fallback only against the local read_runtime_status() record.
_profile_scope swaps process-global skills_tool/skill_manager module
attrs under an RLock; /api/status holds that scope across the
run_in_executor remote-health probe await, so a concurrent
/api/skills?profile=X request can cross-restore the status profile's
skill dir on its finally. Add _config_profile_scope (contextvar-only,
task-local, await-safe) and use it for status, which only resolves
get_hermes_home() at call time for config/env/gateway state and never
needs the skills-module globals.
Context files (AGENTS.md, CLAUDE.md, .hermes.md, .cursorrules, SOUL.md) were
hard-capped at a flat 20K chars before head/tail truncation. Among the agent
harnesses we track, only Codex caps project docs at all (32 KiB); Claude Code,
OpenCode, and Cline load them whole. The flat 20K predates large context
windows and silently truncates real-world AGENTS.md files.
B — dynamic cap: when context_file_max_chars is unset (now the shipped
default), the cap scales with the model's context window
(ctx_tokens * 4 * 0.06, floor 20K, ceiling 500K). Small-context models stay at
the historical 20K; a 200K model gets 48K; large models stop truncating real
docs. An explicit context_file_max_chars still wins. Context length is resolved
once per conversation (stable -> prompt cache untouched).
C — when truncation does happen, the marker now names the concrete file path
and tells the agent to read_file it for the full content.
Validation: 154 targeted tests + full agent/ + hermes_cli/ + test_config
(0 failures); E2E against a real 60K AGENTS.md confirms small windows truncate
with the path-bearing marker, large windows load whole, and the system prompt
is byte-stable across rebuilds.
unicode-bidi:plaintext (#44596) resolves text direction per line, but
list markers and the blockquote border are box chrome driven by the CSS
direction property, which plaintext never sets, so an RTL list renders
its numbers stranded at the far left edge. CSS cannot close this gap
(:dir() only reads the dir attribute, never plaintext resolution), so
ul/ol/blockquote carry dir="auto" and the browser resolves their box
direction natively while the plaintext rules keep owning the text.
Inline code carries dir="ltr", which HTML's auto algorithm skips,
matching the no-vote contract the CSS isolate already gives it.
Rolling back to the oldest curator snapshot failed and deleted that
snapshot. rollback() takes a safety snapshot first, and snapshot_skills()
ends by pruning the backups directory down to keep (5 by default). At the
steady keep limit that prune removed the oldest snapshot, which is the very
one being restored, so the extract found no skills.tar.gz and the rollback
stopped with "snapshot extract failed (state restored)".
Thread an optional protect set through snapshot_skills() into _prune_old()
so the pre rollback safety snapshot can never evict the snapshot being
restored. Add two regression tests covering restore of the oldest snapshot
at the keep limit.
Fixes#47612
The curator now defaults to prune-only: the deterministic inactivity pass
(mark stale / archive long-unused skills) still runs whenever the curator is
enabled, but the opinionated LLM umbrella-building consolidation fork is OFF
by default.
- agent/curator.py: add DEFAULT_CONSOLIDATE=False + get_consolidate(); gate
the forked aux-model review in run_curator_review behind it (new consolidate
param, None=read config). When off, the LLM pass is skipped entirely (no
aux-model cost); the run is still recorded and reported.
- config.py: add curator.consolidate (default false); v29->v30 migration seeds
the key for existing installs without clobbering a user-set value.
- hermes_cli/curator.py: 'hermes curator run --consolidate' override; status
shows consolidate state; prune-only notice on run.
- docs + tests.
When refresh_launchd_plist_if_needed() runs from inside the gateway's own
launchd process tree (agent-initiated self-update via the terminal tool), a
direct launchctl bootout tears down the service's process group — including
the CLI doing the refresh — before the follow-up bootstrap can run. The
gateway is left unloaded and KeepAlive can't revive it (#43842).
Detect in-service execution via gateway.status.get_running_pid() +
_is_pid_ancestor_of_current_process(), and delegate the bootout->bootstrap to
a detached (start_new_session=True) helper that survives the process-group
teardown. The normal out-of-tree CLI path is unchanged.
Fixes#43842.
A turn that ends in an error (e.g. an out-of-funds state) was being
re-rendered in unrelated threads. On a warm thread switch the on-screen
`$messages` still belongs to the previously viewed thread, and
`flushPendingViewState` fed it into `preserveLocalAssistantErrors`, which
grafted the prior thread's failed turn onto the newly opened one. Because
the polluted view then became the next switch's baseline, the error
cascaded into every thread the user visited.
Only carry local errors across a view flush when the on-screen baseline is
the same session being flushed; the cached state we publish already retains
that session's own errors. Also surface the turn error as a global toast
even when the failing turn ran in a background thread, since the error
blocks all subsequent interactions until the user acts.
The double-underscore prefix swap fixed bare native tools but SKIPPED tools
already named mcp_<server>_<tool> (real MCP servers, e.g. mcp_linear_get_issue):
they went on the OAuth wire single-underscore and still tripped Anthropic's
third-party billing classifier -> HTTP 400 'extra usage, not plan limits'.
Verified empirically against a live Max subscription: a single mcp_ tool flips
the whole request to the extra-usage lane; mcp__ is accepted.
- build_anthropic_kwargs: promote ANY leading single-underscore mcp_ to mcp__
(bare names -> mcp__name; mcp_<server>_<tool> -> mcp__<server>_<tool>),
never double-prefixing an already-mcp__ name. Same for tool_use blocks in
history.
- normalize_response: reverse the mcp__ wire name back to whichever original
the registry knows — the single-underscore mcp_<server>_<tool> form for MCP
server tools, or the bare name for native tools — preferring a name that
already resolves natively.
- Tests rewritten to assert the invariant: ZERO single-underscore mcp_ names
reach the OAuth wire, and the mcp__ round-trip resolves back to the
registered name for both native and MCP-server tools.
Builds on liuhao1024's mcp__ prefix commit (cherry-picked). Closes the
MCP-server gap that left any session with an MCP server configured still
billing to extra usage.
Anthropic's Claude-Code request classifier treats tool names with a
single-underscore `mcp_<x>` prefix as non-Claude-Code / third-party,
routing the request to extra-usage billing (HTTP 400). Real Claude Code
uses double underscores: `mcp__<server>__<tool>`.
Change the tool-name prefix from `mcp_` to `mcp__` in both the outgoing
path (build_anthropic_kwargs) and the incoming path
(normalize_response). Update the skip-guard to check for both `mcp_`
and `mcp__` prefixes so native MCP server tools (which use the legacy
single-underscore format) are not double-prefixed.
Fixes#46675
The docstring described a token as path-like when it contains a "/"
separator, but the keystroke-latency fix now excludes "://" scheme tokens
(URLs) even though they contain "/". Document the exclusion so the contract
matches the behavior.
Regression coverage for the keystroke-latency fix: a URL token contains
"/", so the bare-slash path heuristic used to return it as a path word and
run os.listdir on every keystroke. Assert _extract_path_word rejects
http/https/ssh scheme tokens, that ordinary paths (incl. a bare colon) are
unaffected, and that the completer never touches the filesystem for a URL
under the cursor.
The interactive CLI input box runs its completer with
`complete_while_typing=True`, so `SlashCommandCompleter.get_completions`
is invoked on *every* keystroke. That completer does blocking I/O:
fuzzy `@`-file indexing shells out to `rg`/`fd` (up to a 2s timeout) and
file-path completion calls `os.listdir` + `stat`. Because the completer
was passed inline (never wrapped in `ThreadedCompleter`), all of this ran
synchronously on the prompt_toolkit event loop, stalling the render after
each key — very noticeable on WSL2 and other slow-filesystem setups
("typing in the prompt box being very latent").
Two fixes:
- Wrap the input completer in `ThreadedCompleter` so completion work runs
off the UI event loop and never blocks rendering between keystrokes.
- Stop treating URLs as file paths in `_extract_path_word`: a token like
`https://example.com/x` contains `/`, so it triggered `os.listdir` on
every keystroke while typing/pasting a link (listing a bogus `https:`
dir) for a completion that can never be useful. Skip any token with a
`://` scheme separator.
(cherry picked from commit b5be2ba276)
Streamdown runs our `preprocess` inside its own useMemo, and the user
bubble runs `extractEmbeddedImages`/directive parsing inside theirs — so
anything thrown while rendering one message (a regex/stack overflow on
adversarial content) escapes to the ROOT error boundary and takes down
the entire app, as seen in a reported `RangeError: Maximum call stack
size exceeded` from a single message.
Wrap both the assistant preprocess pipeline and the user-message
directive passes in try/catch that degrade to the raw text. One bad
message now renders plain instead of nuking the transcript.
`normalizeFenceBlocks`/`pushProseFence` appended block bodies with
`out.push(...lines)`, which spreads every line as a separate call
argument. A single message carrying a large fenced block (a logged
minified bundle, base64 blob, or big tool dump — common in long
sessions) overflows V8's argument-count limit and throws
`RangeError: Maximum call stack size exceeded`, breaking the transcript
render. Compression doesn't save us: it gates on tokens vs. window, not
a single message's line count, and the protected recent tail renders
verbatim regardless.
Append iteratively via a small `extend()` helper. Behavior is identical
for normal-sized blocks.
Closes#47111
is_container() only recognized Docker (/.dockerenv), Podman
(/run/.containerenv), and docker/podman/lxc markers in /proc/1/cgroup.
Under cgroup v2 (Kubernetes/k3s on containerd or CRI-O) /proc/1/cgroup
collapses to a single "0::/" line with no runtime marker, so
is_container() returned False on every containerd/CRI pod.
That false negative bypassed container-aware behavior across the CLI.
The most damaging case (reported): even after #46290 fixed
detect_service_manager() to gate on _s6_running() alone, other
is_container() call sites (profile home resolution, gateway behaviors,
config, doctor) still misbehave on containerd.
Broaden detection conservatively:
- KUBERNETES_SERVICE_HOST env var (present in every k8s pod).
- kubepods/containerd/crio markers in /proc/1/cgroup (cgroup v1 nested).
- same markers in /proc/self/mountinfo as a cgroup-v2 fallback.
Tests: 3 new (k8s env, kubepods cgroup, cgroup-v2-via-mountinfo) plus the
existing negative case hardened to stub mountinfo + env; 108 constants +
service_manager tests pass.
Follow-up to salvaged PR #41633: the timestamp prefix injection was
unconditional. Gate the in-context render behind
gateway.message_timestamps.enabled (default false) at both the live-message
and history-replay sites; timestamp metadata is still captured + persisted
regardless so the toggle can be flipped on later. Add DEFAULT_CONFIG entry,
docs, and gate tests.
Consolidates these related Amy fork patches:
- 429830f39 feat(gateway): inject message timestamps into user messages for LLM context
- 3c3d6fac0 fix: handle both ISO string and epoch float timestamps in history replay
- 2874f7725 feat: human-friendly timestamp format with weekday and timezone name
- 3735f4c8b fix: render gateway message timestamps once
* fix(desktop): keep the pre-session model pick selected in the picker
The composer picker derived its "current" row from `model.options ?? store`,
so model.options always won. Pre-session that query returns the PROFILE
DEFAULT, not the sticky composer pick — so selecting a model before a session
exists left the checkmark (and the picker's "current" line) on the default,
making the pick look ignored even though the pill updated.
Add `currentPickerSelection()`: with a live session the gateway's model.options
is authoritative; pre-session the sticky `$currentModel`/`$currentProvider`
wins, falling back to options. Wire it into ModelMenuPanel and ModelPickerDialog.
* feat(desktop): global reasoning/speed defaults in Settings → Model
The composer picker is now sticky-UI/per-session only and never writes the
profile default (#46959), but Settings → Model had no reasoning/speed control
and `agent.reasoning_effort` wasn't in the curated config surface at all
(`service_tier` was buried in Advanced) — so there was nowhere to set the
profile default that crons/subagents/messaging resolve from.
Add capability-gated Reasoning (effort) + Fast controls beside the main model,
gated by the applied model's reported capabilities (reasoning defaults on, fast
off when unreported — same as the composer). They read/write `agent.reasoning_effort`
and `agent.service_tier` by round-tripping the config record, matching the
gateway's value semantics (service_tier "fast"/"priority"/"on" ⇒ fast).
* refactor(desktop): don't open the reasoning select from its row label
A <label> wrapping the Select forwarded text clicks to the trigger, opening
the dropdown unexpectedly. Plain row for reasoning; Fast stays a <label> so
clicking its text toggles the switch (expected for a checkbox-like control).
* fix(desktop): re-download Electron binary via mirror when pack fails (#47266)
Since #38673 pinned build.electronDist to node_modules/electron/dist,
electron-builder reads the Electron binary straight from there and never
downloads it during `npm run pack`. That dist tree is only produced by the
electron package's postinstall (install.js) during `npm ci`. When that
download is blocked or throttled (GitHub's release host is unreachable in
some regions), the dist is missing and the build dies with:
The specified electronDist does not exist: .../node_modules/electron/dist
The existing ELECTRON_MIRROR fallback in all three desktop-build paths
(scripts/install.ps1, scripts/install.sh, and `hermes desktop` in
hermes_cli/main.py) re-ran `npm run pack` with ELECTRON_MIRROR set — but
pack never downloads Electron anymore, so the mirror was never used and the
retry re-read the same missing dist. The fallback was effectively dead.
Drive the mirror through electron's own downloader instead:
- Add a dist-presence check + a downloader helper (Test-ElectronDist /
Restore-ElectronDist, _electron_dist_ok / _restore_electron_dist,
_electron_dist_ok / _redownload_electron_dist) that wipes a partial dist
+ the path.txt version marker (electron's install.js short-circuits on it)
and re-runs `node install.js`, optionally via a mirror.
- On the first retry, repopulate a missing dist from the canonical source;
on the mirror retry, re-fetch through npmmirror.com, then pack.
- Gate the re-download on the dist check so an unrelated build failure
(tsc/vite) doesn't trigger a pointless ~200 MB refetch, and skip the final
pack when the binary still can't be fetched instead of failing the same way.
* test(desktop): cover Electron dist re-download mirror fallback (#47266)
Add behavior coverage for the electronDist re-download fix:
- _electron_dist_ok across linux/win32/darwin, including the partial-dist
case (dir present but binary missing) that makes the pinned electronDist
fail.
- _redownload_electron_dist: no-op when the binary is present, bail when
install.js is absent, wipe a stale dist + path.txt marker and run
electron's downloader with ELECTRON_MIRROR injected, and report failure
when the download still produces no binary.
- `hermes desktop`: the mirror fallback now drives electron's own downloader
before re-running pack, and skips the final pack entirely when the binary
can't be fetched.
Replaces the old mirror test that asserted the (now-fixed) dead behavior of
re-running `npm run pack` with ELECTRON_MIRROR set — pack never downloads
Electron under the pinned electronDist, so that retry could never help.
Guards that two user-defined custom endpoints exposing an overlapping
model each keep their full catalog — the dedup must never cross-filter
two user-defined rows against each other.
The /model interactive picker resolved a base_url from user credentials
but never passed it to ProviderProfile.fetch_models(), causing the
picker to always query the provider's hardcoded default endpoint
instead of the user's custom URL (e.g. a company litellm proxy).
- providers/base.py: add optional base_url parameter to fetch_models()
- hermes_cli/models.py: pass resolved base_url to fetch_models()
- Update all subclass overrides for signature compatibility
- Add 6 regression tests covering override, fallback, and integration
Support files under references/, templates/, assets/, and scripts/ are progressive-disclosure data loaded through skill_view(..., file_path=...). They should not be treated as standalone skills during discovery or collision checks.
This prevents archived skill packages or support markdown files inside a real skill from shadowing active skills with the same name while still allowing top-level categories named scripts/templates/assets/references.
Tests cover:
- pruning nested SKILL.md files inside skill support directories
- preserving support-named top-level categories
- avoiding skill_view collisions from support markdown
- keeping archived package SKILL.md files accessible only through file_path
Salvage follow-up for PR #29575: add regression tests for the section-3
no-api_key /v1/models probe (probes bare endpoints, skips when explicit
models set) and add the contributor AUTHOR_MAP entry.
Section 3 of list_authenticated_providers (user-defined endpoints from
the providers: config section) required an api_key before probing the
endpoint's /v1/models for live model discovery. This broke local
self-hosted backends (llama.cpp, Ollama, vLLM, etc.) that don't require
authentication — they would only ever show the single default_model
from config instead of the full model catalog.
Section 4 (custom_providers list) already handled this correctly with
the policy: probe when api_key is set OR when no explicit models are
configured. Apply the same logic to Section 3 so local backends get
full model discovery without requiring a placeholder api_key workaround.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cover #47375 fix: record-on-rich-send + lookup-on-reply round trip,
lookup miss leaving reply_to_text None, and precedence (native quote
and echoed caption both win over the index fallback).
Telegram does not echo a sendRichMessage's content back in
reply_to_message (.text/.caption empty, .api_kwargs None), so replies
to rich sends (briefings, the gateway's own rich finals) arrived with
no quotable text and the [Replying to: ...] injection was skipped.
Remember message_id -> text at send time in a best-effort JSON index
(gateway/rich_sent_store.py), and recover it on inbound when text and
caption are both empty. Best-effort and no-throw throughout: any
failure degrades to prior behavior and never breaks a send or message.
Salvaged from #47375 by @x1erra. Dropped the cross-platform run.py
reply-prefix rewrite (out of scope; bloated every reply on every
platform) and scrubbed a docstring reference to an out-of-repo script.
Kept the inbound reply_to logging enrichment used to verify the fix.
The #45954 model-dedup builds `user_models` from every is_user_defined
row, then strips those model IDs from every row where is_aggregator(slug)
is True. But is_aggregator() returns True for *every* `custom:*` slug, and
list_authenticated_providers emits named custom providers with slug
`custom:<name>` and is_user_defined=True. So a user's own custom provider
is treated as an aggregator and filtered against user_models — which holds
exactly its own models (the row helped build that set). Every model is
removed, the row drops to zero, and the provider disappears from the model
picker.
Guard the dedup loop to skip is_user_defined rows: a user's configured
provider is never an aggregator duplicate of itself. Built-in aggregators
(openrouter, etc.) are still deduped as before. Adds a regression test.
Reflect the default-model change in the xAI Grok OAuth guide, the web
search docs (EN + zh-Hans), and the web provider docstring. grok-4.3 is
kept in the model tables as the previous default; the Nous/OpenRouter
aggregator catalog still lists grok-4.3 and is left unchanged.
Switch the default model for the xAI/Grok provider and the xAI web
search backend from grok-4.3 to grok-build-0.1. grok-build-0.1 is
already recognized by the model metadata, so no new model definition
is required; grok-4.3 remains selectable.
Follow-up to salvaged PR #41624:
- Remove stray urllib.parse import in run_agent.py (cherry-pick cruft, unused)
- Add tests: session:compress emits with correct context, no-callback is
safe, and a callback exception does not break compression
* feat(desktop): stream subagent replies into watch windows
A desktop watch window resumes a child session lazily (no full agent) and
mirrors the parent-relayed `subagent.*` events into native child-session
stream events. The child's streamed reply text was never relayed, so the
window sat blank while the subagent "talked".
- delegate_tool: forward the child's `run_conversation` stream tokens up the
progress relay as `subagent.text` (inert under CLI/TUI — their progress
handlers ignore non-tool event types; only a gateway watch window mirrors it).
- server: mirror `subagent.text` -> `message.delta` on the child sid only, and
skip the parent emit (per-token frames are meaningless on the parent session,
which shows the child via the spawn tree). Demote `subagent.start` to a
one-time goal header and drop the noisy `subagent.progress` mirror — tools
already mirror natively.
- server: guard `_start_agent_build` so a lazy watch session spectating an
in-flight child stays lazy; incidental RPCs were upgrading it to a full
agent mid-stream and silently killing the mirror.
* fix(desktop): keep watch-window chat clear of titlebar chrome
Secondary windows (new-session scratch, subagent watch, cmd-click pop-out)
hide the titlebar tool cluster + session header, so the transcript ran to the
window's top edge and streamed text slid up under the OS traffic lights.
- Gate the hidden chrome on `isSecondaryWindow()` everywhere (app-shell,
chat header, thread list) instead of the narrower new-session flag.
- Add a fixed opaque drag-strip at the top of the secondary-window transcript:
content padding alone scrolls away with the text, so the strip masks
anything behind it and keeps the window draggable like the main header.
* fix: WSL subagent window
* fix: subagent window top padding
---------
Co-authored-by: Austin Pickett <pickett.austin@gmail.com>
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
Follow-up to salvaged PR #41619: replace the module-global
_truncation_warnings list with a contextvars.ContextVar so concurrent
gateway-session prompt builds can't drain or clear each other's pending
warnings (cross-session leak). Adds a context-isolation test.
PROBLEM: Automatic context files such as SOUL.md and AGENTS.md were capped by a hardcoded CONTEXT_FILE_MAX_CHARS value. Amy's local fork had raised that constant from 20K to 25K so a larger SOUL.md would not be silently truncated, but the hardcoded 25K value changed upstream default behavior and made the patch less generally useful.
SOLUTION: Restore the upstream-compatible 20K default, add a context_file_max_chars config setting for users who intentionally keep larger identity/project-context files, keep chat-visible truncation warnings, and document the new setting. Tests cover the default, config override, explicit max_chars precedence, and the warning text.
