wfb 0.1.4

Give terminal agents persistent memory, Gemini OAuth sessions, and live Chrome context with a zero-dependency local CLI

wayfarer-bridge

Wayfarer Bridge (wfb) is a local, Python standard library-only CLI for moving useful context between terminal agents and Gemini. It is intentionally small: local files, SQLite, OAuth, and raw Gemini REST calls with no external Python package dependencies.

The project currently has two working context channels:

• World state: a local SQLite store for structured tasks, constraints, and style rules (wfb seed, wfb status).
• Gemini chat memory: local Gemini session files that make repeated wfb gemini ask calls coherent across turns (wfb gemini session, wfb gemini ask).

Direct attachment to active browser Gemini side-panel sessions is not available through the validated official Gemini API surface today. For now, wfb uses local session memory as the bridgeable, agent-friendly fallback.

Installed from PyPI or any environment: The ordered workflow (init, doctor, bridge vs chrome commands, multi-tab rules) is printed at the end of wfb --help under Agent workflow guidance, so you get the same playbook from a development alias, a global pip install wfb, or CI. This README and docs/chrome_remote_debug_bridge.md add Chrome CDP detail.

For terminal agents (Chrome CDP)

• Debug instance only: wfb only sees tabs on the Chrome process that exposes the remote-debug HTTP endpoint for the port and profile you use. A normal daily browser window without remote debugging does not appear in wfb chrome targets.
• Blind or recovery start: Run wfb bridge doctor --format json for endpoint reachability, requested vs resolved port, attachment state, and suggested next commands.
• Port probing: wfb chrome capture, wfb chrome inspect, wfb chrome ax, wfb chrome find, and wfb bridge … try the requested port first, then other detected debug ports. wfb chrome targets and wfb chrome attach use --port only (no automatic fallback). Use the port from doctor or pass --port explicitly when listing or attaching.
• Goals → commands: For Gemini plus live page context with AOM auto-selection, use wfb bridge ask / wfb bridge loop (default --capture-mode auto). For read-only structured context without Gemini, wfb chrome attach --target-id … then wfb chrome ax / wfb chrome find. For a quick text-only snapshot, wfb chrome capture — it re-resolves which tab to use on every invocation (bounded text via inspect, not the AX tree).
• Stable tab vs blind capture: wfb chrome inspect, ax, and find follow the persisted attachment when you omit --target-id. wfb chrome capture and wfb bridge ask/loop call target selection each time; with multiple tabs they may pick a different tab unless you pass --target-id or attach first and use inspect/ax/find.
• Multi-tab rule: If chrome targets lists more than one candidate, do not rely on blind capture/bridge ask without checking JSON warnings and selection.method, or pass --target-id. JSON may warn when the tab was chosen by heuristic and Chrome did not report an active target.
• Metadata vs snapshot: chrome targets and doctor list ids, titles, URLs, and types. capture is text-only; inspect, ax, find, and bridge supply bounded text and/or accessibility data. Sensitive URLs still show up in listings.
• AOM on structured pages: Prefer wfb chrome ax --format outline (after attaching to the right tab) over raising --max-chars on chrome inspect. The accessibility tree is screen-reader-style and dramatically more token-efficient than body.innerText for chat logs, forms, and component-heavy UIs. Use wfb chrome find --query "..." to locate text without re-fetching ever-larger snapshots.
• Auto capture mode: wfb bridge ask and wfb bridge loop default to --capture-mode auto, which inspects the AX tree, picks AOM when the page has meaningful structure, and falls back to text otherwise. Override with --capture-mode text|aom|both when you want a fixed strategy.
• Internal Chrome URLs: chrome://… pages often produce empty or unhelpful snapshots compared with normal https:// tabs. When several targets exist (for example Google AI internal UI vs the search results page), prefer --target-id from chrome targets.

What Works Today

• wfb init creates the local asset directory, initializes the SQLite DB, and runs OAuth login.
• wfb seed ingests a structured world-state JSON envelope into SQLite.
• wfb status prints a concise world-state summary in text or JSON.
• wfb gemini ping verifies authenticated Gemini API access.
• wfb gemini ask sends prompts to Gemini with local session continuity.
• wfb gemini session ... creates, switches, lists, resets, and inspects local chat sessions.
• wfb gemini ask --auto-summarize on can compact long local sessions with a Gemini-generated summary.
• wfb chrome ... can launch a debuggable Chrome instance, list page targets, attach one target, inspect bounded tab text context, capture the page accessibility tree (chrome ax), and search the page text and AX tree (chrome find).

Still not supported:

• Direct official Gemini API attachment to browser side-panel sessions.
• A centralized OAuth client secret for the open-source/PyPI path.

Quick Start

Install locally (editable) from the repository root:

python3 -m pip install -e .

Then use:

wfb --help

You can also run the script directly:

python3 wfb.py --help

For isolated testing, install in a virtual environment first.

Set up OAuth:

1. Create a Google Cloud project and enable the Generative Language API.
2. Configure the OAuth consent screen and add yourself as a test user while developing.
3. Create an OAuth Desktop app client.
4. Download the JSON and place it at ~/.wfb/client_secret.json.
5. Run wfb init, then wfb gemini ping.

Create a Gemini session and ask follow-up questions:

wfb gemini session new --name planning
wfb gemini ask --prompt "Remember that Wayfarer Bridge is stdlib-only."
wfb gemini ask --prompt "What constraint did I just mention?"
wfb gemini session inspect --format json

Asset Layout

~/.wfb/ is the canonical per-user asset directory.

Path	Purpose
~/.wfb/wayfarer.db	Default SQLite world-state database.
~/.wfb/client_secret.json	User-provided OAuth Desktop client secret.
~/.wfb/token.json	Local OAuth token cache.
~/.wfb/gemini_sessions/<session_id>.json	Local Gemini chat session history and metadata.
~/.wfb/gemini_active_session.json	Pointer to the current active Gemini session.
~/.wfb/chrome_debug_profile/	Isolated Chrome profile dir for wfb chrome launch --profile-mode isolated.
~/.wfb/chrome_attachment.json	Persisted selected Chrome target for wfb chrome inspect.

The global --db PATH flag overrides only the SQLite database path. Other CLI assets still live under ~/.wfb/.

CLI Reference

Global option:

wfb [--db PATH] <command> ...

wfb init [--no-open-oauth-guide] [--no-browser] [--force-login]

• Ensures ~/.wfb/ exists.
• Initializes the SQLite schema at --db PATH or ~/.wfb/wayfarer.db.
• Requires ~/.wfb/client_secret.json.
• Runs OAuth login and stores token data at ~/.wfb/token.json.
• --no-open-oauth-guide disables best-effort browser-open for setup instructions.
• --no-browser prints the OAuth URL instead of attempting to open a browser.
• --force-login ignores a valid cached token and reruns OAuth.

wfb seed (--json STRING | --file PATH) [--replace]

Ingests a structured world-state envelope into SQLite. Default behavior is upsert by id; --replace clears all entity tables before inserting the supplied envelope.

wfb status [--format text|json] [--limit N]

Prints the SQLite world-state summary. --format json emits deterministic machine-readable state; --limit N caps preview lists in text output.

wfb gemini ping [--limit N]

Lists available Gemini model names using cached OAuth credentials.

wfb gemini ask --prompt STRING [options]

Sends a prompt to Gemini using local session memory.

Options:

• --model ID: Gemini model id. Default: gemini-2.5-flash.
• --session ID: route this ask to a specific local session. Defaults to the active session.
• --max-history-turns N: include at most this many recent non-summary turns. Default: 30.
• --system TEXT: system instruction override for this call.
• --auto-summarize on|off: opt into model-generated session compaction. Default: off.
• --summarize-model ID: optional model override used only for summary generation.
• --sync-world-state on|off: per-ask override for world-state sync.
• --world-state-db PATH: per-ask override for sync target DB.

If no active session exists, ask auto-creates one.

wfb gemini session current

Prints the active local Gemini session id, if any.

wfb gemini session list

Lists local sessions. The active session is marked with *.

wfb gemini session new [--name NAME] [--model ID] [--system TEXT] [sync options]

Creates a new local session and makes it active.

Sync options:

• --sync-world-state on|off: default sync mode for this session.
• --world-state-db PATH: default target DB for this session's sync.
• --world-state-scope TEXT: optional scope tag injected into synced record metadata.

wfb gemini session use --id SESSION_ID [sync options]

Makes an existing session active.

wfb gemini session reset [--id SESSION_ID]

Clears message history for the target session. Defaults to the active session.

wfb gemini session inspect [--id SESSION_ID] [--format text|json]

Inspects a local session. Defaults to the active session.

wfb chrome launch [--port N] [--profile-mode isolated|user] [--chrome-path PATH]

Launches a Chrome instance with remote debugging enabled and verifies /json/version. If a debug endpoint is already available on the port, wfb reuses it and skips spawning a new process. If the requested port is unavailable, wfb auto-falls back to a detected healthy debug port.

• --profile-mode isolated (default): uses ~/.wfb/chrome_debug_profile/.
• --profile-mode user: uses your normal Chrome profile (higher fidelity, higher risk).

wfb chrome targets [--port N] [--include-types page,webview] [--gemini-only] [--format text|json]

Lists attachable CDP targets from /json/list.

• Default target types: page (backward-compatible behavior).
• --include-types page,webview includes Gemini side-panel webviews.
• --gemini-only narrows output to Gemini-related targets.

wfb chrome attach --target-id ID [--port N] [--include-types page,webview] [--format text|json]

Stores the selected target in ~/.wfb/chrome_attachment.json for later inspection.

• Default target types searched: page.
• Use --include-types page,webview when attaching Gemini side-panel IDs.

wfb chrome inspect [--target-id ID] [--port N] [--include-types page,webview] [--max-chars N] [--selector CSS] [--format text|json]

Reads bounded tab context from CDP Runtime.evaluate and prints JSON by default:

• url, title
• selected_text
• text_snapshot (bounded/truncated exactly by --max-chars)
• capture metadata (captured_at_unix, length fields, target metadata)
• selector, selector_matched when --selector CSS is supplied

Inspect defaults are attachment-aware: when using a persisted attachment and no explicit --include-types is provided, wfb auto-includes the attachment target type (for example webview) so inspect can resolve saved side-panel targets without extra flags.

--selector CSS evaluates document.querySelector(selector) and returns only that element's innerText. When the selector matches nothing, JSON output reports selector_matched: false, the snapshot is empty, and a recovery hint is written to stderr instead of silently falling back to body.innerText.

Debug-port probing: inspect tries your --port (default 9222) first, then other locally detected debugging ports in stable order until /json/list succeeds — same ordering as chrome capture/bridge. When JSON output is enabled, a top-level debug object documents requested_port, resolved_port, and whether a fallback port was used (fallback_used).

wfb chrome ax [--target-id ID] [--port N] [--include-types page,webview] [--role ROLE] [--name SUBSTRING] [--depth N] [--max-nodes N] [--name-max-chars N] [--ignored on|off] [--timeout-seconds F] [--format outline|json]

Reads the page Accessibility Tree using CDP Accessibility.getFullAXTree and renders a screen-reader-style outline. Designed to replace the "raise --max-chars and re-run" pattern on structured pages: outlines are typically an order of magnitude smaller than the equivalent body.innerText snapshot.

Outline format example:

WebArea
  main "Conversation"
    log "Messages"
      paragraph "Hi there"
      paragraph "How can I help?"
    textbox "Compose" focused level=1
    button "Send" disabled

• --role ROLE and --name SUBSTRING narrow the rendered outline to subtrees rooted at matching nodes. Useful for "show me only the main landmark" or "only the log region" without scanning the entire page.
• --depth N is forwarded to CDP as a source-side bound on tree size.
• --max-nodes N (default 600) caps rendered nodes; --name-max-chars N (default 120) truncates very long accessible names with a "…(+N chars)" indicator.
• --ignored on includes AX nodes flagged as ignored (default skips them and re-emits their non-ignored children at the parent depth, matching screen-reader behavior).
• JSON output includes outline_meta, ax_quality (meaningful_roles, generic_roles, meaningful_ratio), filter echo, normalized nodes (with backend_dom_node_id for future write-path commands), and debug.

Example:

wfb chrome ax --format outline
wfb chrome ax --format json --role textbox
wfb chrome ax --format outline --name "send" --max-nodes 80

wfb chrome find --query STRING [--mode text|aom|both] [--target-id ID] [--port N] [--include-types page,webview] [--selector CSS] [--role ROLE] [--max-results N] [--context-chars N] [--max-chars N] [--ax-depth N] [--timeout-seconds F] [--format text|json]

Searches the page once and returns matches with surrounding context.

• --mode text searches the page text snapshot (or selector subtree text). Each match includes before, match, after, and the byte offset.
• --mode aom searches AX node names (case-insensitive) and returns each match plus a role/name breadcrumb path from the AX root.
• --mode both (default) does both and returns each list separately.
• --selector CSS scopes text-mode search to a subtree.
• --role ROLE restricts AOM matches to a specific role.

Designed to replace the "raise --max-chars and retry" workflow: instead of repeatedly re-fetching larger snapshots, ask exactly the question you want to answer ("does the page mention error?", "is there a button named Send?").

Example:

wfb chrome find --query "session expired" --mode both
wfb chrome find --query "Send" --mode aom --role button --format json
wfb chrome find --query "TODO" --mode text --selector "main" --context-chars 60

wfb chrome detach

Clears the persisted attachment file.

wfb chrome current [--format json|text]

Shows the current persisted attachment and endpoint health. Default output is JSON.

wfb chrome capture [--target-id ID] [--port N] [--include-types page,webview] [--gemini-only] [--max-chars N] [--selector CSS] [--format json|text]

Runs discover -> attach -> inspect in one deterministic command.

• Defaults to --include-types page,webview for side-panel compatibility.
• Selection priority: explicit --target-id, then focused/active target, then heuristic ranking, then first-candidate fallback.
• --selector CSS scopes text snapshot to a subtree (same semantics as chrome inspect).
• JSON output includes selection, target, attachment, inspect, and debug ports (requested_port, resolved_port).

wfb bridge doctor [--port N] [--include-types page,webview] [--gemini-only] [--format json|text]

Read-only diagnostics for agents: endpoint reachability (/json/version), target summaries (/json/list), persisted attachment state, Gemini session id, and ordered recommendations (recover / happy-path next commands). Default format is JSON; --format text prints a compact checklist.

Example:

wfb bridge doctor
wfb bridge doctor --include-types page,webview --gemini-only --format text

wfb bridge ask --prompt "..." [--session ID] [--model MODEL] [--system TEXT] [--port N] [--target-id ID] [--include-types page,webview] [--gemini-only] [--max-chars N] [--selector CSS] [--capture-mode text|aom|both|auto] [--ax-max-nodes N] [--ax-name-max-chars N] [--format json|text]

Runs the full browser-to-Gemini pipeline in one command: capture -> prompt envelope -> Gemini ask.

• Captures browser context via the same logic as chrome capture/chrome ax (including unified debug-port probing: requested_port vs resolved_port in the capture section when fallback occurs).
• --capture-mode auto (default) inspects the AX tree, picks AOM when the page has at least 5 meaningful AX roles and a meaningful_ratio >= 0.3, and falls back to text otherwise. Override with text, aom, or both for a fixed strategy.
• --selector CSS scopes text snapshot to a subtree (same semantics as chrome inspect).
• --ax-max-nodes / --ax-name-max-chars bound AOM outline size when capture-mode includes AOM.
• Builds a versioned prompt envelope (template "2") embedding the chosen capture content and user prompt.
• Sends the composed prompt to Gemini within the target session (creates one if none active).
• Appends both the composed prompt and model response to session history.

JSON output includes three provenance sections:

• capture: target selection, snapshot metadata, selector/selector_matched, mode_requested/mode_chosen/mode_reason, ax_quality, debug ports.
• prompt_envelope: template version, original user prompt, composed prompt character count, and budget (chosen capture mode, text snapshot chars/truncation, AX total/rendered nodes, outline truncation).
• gemini_response: model, session id, full answer text.

Example:

wfb bridge ask --prompt "summarize the feedback on this page" --format json
wfb bridge ask --prompt "extract the test failures" --capture-mode aom --format text
wfb bridge ask --prompt "what's in the main article?" --selector "main article"

wfb bridge loop --prompt "..." [--max-iterations N] [--stability-check on|off] [--session ID] [--model MODEL] [--system TEXT] [--port N] [--target-id ID] [--include-types page,webview] [--gemini-only] [--max-chars N] [--selector CSS] [--capture-mode text|aom|both|auto] [--ax-max-nodes N] [--ax-name-max-chars N] [--format json|text]

Runs a bounded iterative automation loop: each iteration captures browser context and asks Gemini about it.

• Uses the same capture path as wfb bridge ask for port probing, capture-mode selection, and provenance (requested_port/resolved_port per iteration).
• --max-iterations N (default 3): hard upper bound on iterations.
• --stability-check on: stops early if the captured content (text snapshot and/or AX outline depending on capture mode) is unchanged between iterations.
• Each iteration produces full provenance in the output transcript, including the per-iteration prompt_envelope.budget.

Stop reasons:

• max_iterations: budget exhausted normally.
• no_change: stability check detected identical content.
• error: a capture or ask stage failed (error details in the iteration record).

JSON output includes:

• run: start/end timestamps, iteration budget, stop reason.
• iterations[]: per-step capture, prompt_envelope, gemini_response, and status.
• summary: last answer, session id, model, and stop reason.

Example:

wfb bridge loop --prompt "check for new test results" --max-iterations 5 --stability-check on
wfb bridge loop --prompt "summarize changes" --capture-mode aom --format text

Gemini Sessions For Agents

The Gemini REST calls used by wfb do not return a reusable API-managed conversation handle. See docs/gemini_session_discovery.md for the discovery notes.

wfb therefore maintains local session history:

flowchart TD
    askCmd[geminiAsk] --> loadSession[loadLocalSession]
    loadSession --> buildPrompt[buildPromptHistory]
    buildPrompt --> geminiApi[GeminiGenerateContent]
    geminiApi --> appendTurns[appendUserAndModelTurns]
    appendTurns --> saveSession[saveLocalSession]

Session behavior:

• wfb gemini ask targets the active session by default.
• --session ID overrides routing for one call and makes that session active.
• Each successful ask appends both the user prompt and Gemini response.
• session inspect --format json exposes the stored history for orchestration.
• session reset clears a session without deleting the session file.

Chat To World-State Sync

wfb can distill chat context into the SQLite world-state store after successful asks.

Mode and source of truth:

• Chat is primary context.
• World state is extracted, structured output.
• Sync is session-default configurable and per-ask overridable.

Enable sync by default for a session:

wfb gemini session new --name triage --sync-world-state on

Or update an existing session:

wfb gemini session use --id sess_abc --sync-world-state on --world-state-db ~/.wfb/wayfarer.db --world-state-scope triage

Override for a single ask:

wfb gemini ask --prompt "..." --sync-world-state on --world-state-db ~/.wfb/wayfarer.db

Pipeline:

1. ask gets model response and persists chat turns.
2. If sync is enabled, wfb asks Gemini to emit a strict v1 seed envelope JSON.
3. Envelope is validated with the same seed validators.
4. Valid rows are upserted into the target DB via seed_db(...).

Failure semantics:

• Sync failures are non-fatal for chat continuity.
• Ask still succeeds and prints response.
• Sync failures are emitted as deterministic warnings to stderr.

Session Summarization

Summarization is opt-in:

wfb gemini ask --auto-summarize on --prompt "continue"

When enabled and a session exceeds model-aware drift thresholds, wfb asks Gemini to summarize older turns. It then stores a synthetic history_summary message plus recent raw turns.

Important safety details:

• Summaries are Gemini-generated only; there is no deterministic fallback.
• If summary generation fails, ask fails and reports the API error.
• Compacted state is persisted only after the final ask succeeds, so transient final-call failures do not overwrite raw history.
• Summary artifacts are always included in prompt history even when --max-history-turns trims older ordinary turns.
• Thresholds are drift heuristics, not hard Gemini context-window limits.

OAuth Setup

The open-source wfb CLI does not ship a centralized OAuth client secret. Users provide their own Desktop OAuth client JSON at:

~/.wfb/client_secret.json

Reference:

• Gemini OAuth quickstart

Troubleshooting:

• If browser-open fails, copy/paste the printed auth URL manually.
• For headless/manual environments, run wfb init --no-browser.
• If token refresh fails, rerun wfb init --force-login.
• In testing-mode OAuth projects, refresh tokens may expire periodically and require re-login.

World-State Reference

The SQLite world-state store is the original relational context channel. It is useful for durable, structured facts that agents should be able to query or summarize without replaying chat history.

Seed Envelope

Top-level JSON object:

Field	Required	Notes
version	Yes	Integer; must equal 1.
generated_at	No	ISO-8601 UTC string recommended.
source	No	Short origin label, e.g. gemini or cursor.
active_tasks	No	Array; omit or [] for none.
environmental_constraints	No	Array; omit or [] for none.
style_specifications	No	Array; omit or [] for none.

Example:

{
  "version": 1,
  "generated_at": "2026-05-05T23:00:00Z",
  "source": "gemini",
  "active_tasks": [],
  "environmental_constraints": [],
  "style_specifications": []
}

Record Shapes

Validation runs before any DB writes. metadata, when present, must be a JSON object and is stored as compact JSON text in metadata_json.

active_tasks[]

Field	Required	Type	Constraints
id	Yes	string	Primary upsert key.
title	Yes	string
status	Yes	string	One of pending, in_progress, blocked, done.
priority	No	int	Default 0.
owner	No	string
due_at	No	string	ISO-8601 recommended.
notes	No	string
source	No	string	Overrides envelope source.
metadata	No	object

environmental_constraints[]

Field	Required	Type	Constraints
id	Yes	string	Primary upsert key.
kind	Yes	string	One of tool_version_warning, policy, runtime_limit, dependency, other.
name	Yes	string
value	Yes	string
severity	Yes	string	One of info, warn, error.
scope	No	string	e.g. global, repo, task:<id>.
source	No	string	Overrides envelope source.
metadata	No	object

style_specifications[]

Field	Required	Type	Constraints
id	Yes	string	Primary upsert key.
category	Yes	string	One of tone, formatting, coding_style, workflow, other.
rule	Yes	string
priority	No	int	Default 0.
applies_to	No	string	e.g. all, python, docs.
source	No	string	Overrides envelope source.
metadata	No	object

Upsert Semantics

• Rows are inserted or updated by id.
• updated_at is refreshed on every successful write.
• Row source uses item source, then envelope source, then NULL.
• Unknown top-level envelope keys are rejected.
• Unknown keys inside entity records are rejected.

SQLite Schema

wfb init creates a v1 schema with one schema_version row and three entity tables:

• active_tasks
• environmental_constraints
• style_specifications

Indexes:

• idx_active_tasks_status
• idx_constraints_severity
• idx_style_priority

The implementation source of truth is wfb_db.py.

wfb status --format json

Top-level shape:

{
  "version": 1,
  "db_path": "/Users/you/.wfb/wayfarer.db",
  "summary": {
    "tasks": {
      "pending": 0,
      "in_progress": 0,
      "blocked": 0,
      "done": 0
    },
    "constraints": {
      "info": 0,
      "warn": 0,
      "error": 0
    },
    "style_specifications": 0
  },
  "highlights": {
    "tasks": [],
    "constraints": [],
    "style_specifications": []
  },
  "updated_at": {
    "active_tasks": null,
    "environmental_constraints": null,
    "style_specifications": null
  }
}

db_path is the resolved absolute database path. highlights.* rows use SQL column names; metadata_json remains a JSON string.

Exit Codes

Code	Meaning
0	Success.
2	CLI usage / argument error.
3	Validation error.
4	Database error.
5	File I/O, OAuth, token refresh, network, or Gemini API error.

Implementation Notes

• Python standard library only.
• No external Python package requirements.
• wfb.py is the CLI entrypoint.
• Supporting modules:
  • wfb_paths.py: local asset paths.
  • wfb_db.py: SQLite schema and lifecycle.
  • wfb_oauth.py: OAuth installed-app flow and token storage.
  • wfb_gemini_api.py: Gemini REST client, model calls, summarization.
  • wfb_gemini_sessions.py: local session storage and compaction helpers.
  • wfb_chrome_bridge.py: stdlib CDP discovery, websocket, and page inspection helpers.
  • wfb_chrome_session.py: persisted Chrome target attachment record helpers.

The security posture is intentionally conservative: local-first, stdlib-only, no shared OSS OAuth client secret, and no vendored third-party libraries yet. A deeper security review is still future work.

Known Limits / Next Work

• No direct browser Gemini session attach through official API endpoints validated so far.
• Chrome bridge is currently macOS-focused for launch ergonomics.
• wfb chrome inspect is read-only text extraction only (no click/type/navigation automation).
• No retry/backoff policy for transient Gemini 429 / 503 errors yet.
• No tool/function-calling support yet.
• Summarization thresholds are heuristic and should be adjusted with real usage.
• Planned next CLI ergonomics after webview support include wrapper capture flow and trust-profile work.