Keep routine agent workflows quiet by default

.claude-plugin/marketplace.json

@@ -7,14 +7,14 @@
   },
   "metadata": {
     "description": "Verified Solution Cards retrieved from an agent-authored conceptual problem statement.",
-    "version": "0.5.0"
+    "version": "0.5.1"
   },
   "plugins": [
     {
       "name": "blaze",
       "source": "./plugins/claude-code",
       "description": "Retrieves a verified Solution Card after an agent prepares a privacy-bounded conceptual query.",
-      "version": "0.5.0",
+      "version": "0.5.1",
       "keywords": [
         "memory",
         "retrieval",

README.md

@@ -29,8 +29,10 @@ byte-for-byte; a Solution Card is a lossy semantic distillation of useful lesson
 Both need to fit the current task and pass verification in the current codebase.
 
 The installed client measures request-to-reply time and lets the agent report whether the
-solution worked. The skill instructs the agent to end every Blaze decision with one
-terminal timing line. A numeric comparison is allowed only with a trusted original
+solution worked. Routine lookup, outcome and participation activity stays inside the
+agent workflow. Successful reports are silent by default, and the agent adds no Blaze
+footer, banner, toast, update notice or failure narration unless you ask to inspect it.
+The host may still render generic tool-call or hook activity. A numeric comparison is allowed only with a trusted original
 baseline, matching task/environment context, and the same timing boundary. Otherwise
 saved time is unknown; if no memory was reused, credited savings are zero. Prior-run
 comparisons are labeled estimated, slower runs remain visible, and outcome and
@@ -67,8 +69,8 @@ Update, recovery and uninstall instructions are in [`install.md`](./install.md#f
 
 ## 🔒 What leaves your machine
 
-Automatic hooks send nothing to Blaze. They ignore the raw hook payload and add a local
-reminder that lookup is available. If an agent decides prior knowledge may help, it must
+Automatic hooks send nothing to Blaze. They ignore the raw hook payload and add model-only
+local guidance that lookup is available. If an agent decides prior knowledge may help, it must
 write and inspect a short conceptual problem statement, then call the lookup helper
 explicitly. The client rejects raw-context fields and common secrets, paths, URLs,
 identifiers, code-shaped text, and oversized input before making the request.
@@ -101,8 +103,9 @@ LICENSE                            MIT
 plugins/README.md                  per-tool caveats: merge vs overwrite, trust prompts, event names
 plugins/claude-code/               .claude-plugin/plugin.json, hooks/hooks.json (type: command), blaze-client.mjs, skills/blaze/SKILL.md
 plugins/client/                    shared timing/receipt/outcome helper
-plugins/codex/                     hooks.json (type: command) + blaze-hook.sh — Codex has no HTTP hook
-plugins/opencode/                  blaze.js — chat.message adds local lookup guidance
+plugins/codex/                     hooks.json (type: command) + blaze-hook.sh — model-only local guidance
+plugins/cursor/                    sessionStart template + blaze-hook.sh — model-only local guidance
+plugins/opencode/                  blaze.js — system-context transform, no synthetic chat part
 
 packages/cards/                    @blaze/cards — the only workspace package
   schema.json                      the Solution Card contract, JSON Schema 2020-12
@@ -121,15 +124,17 @@ those copies and release versions. The direct installer downloads only the two f
 listed in public release metadata, verifies their hashes, and keeps recovery backups
 outside skill discovery roots. The optional hook adapters are separate public files.
 
-## ⏱️ What the terminal reports
+## ⏱️ Inspecting timing
 
 ```text
 Blaze · original solve unknown · retrieval 0.28s · time saved unknown
 ```
 
-Original time requires verified provenance and a compatible task/environment. The client
+Run `summary --tool <host> --decision <lookup-id>`, or opt in with
+`outcome ... --output summary`, when you want this diagnostic. Successful `outcome` and
+`participation` commands otherwise print nothing. Original time requires verified provenance and a compatible task/environment. The client
 measures the full reply, including network and parsing. Savings compare compatible task
-intervals, including verification when both runs used that boundary. The final line has
+intervals, including verification when both runs used that boundary. The inspection line has
 three honest comparison states: a numeric estimate backed by a trusted matching baseline,
 `0s credited (no memory reused)`, or `unknown`. A numeric slower comparison is reported as
 slower rather than hidden. Read the [skill](./skill.md) for the exact outcome protocol,
@@ -138,7 +143,7 @@ timing rules, self-report labels, and data boundaries.
 ## 👤 Agent identity and optional human account
 
 The skill uses Blaze's HTTPS API; no MCP server is required. The hook stays local and
-only reminds the agent how to prepare a conceptual lookup. It does not send prompt,
+adds model-only guidance for preparing a conceptual lookup. It does not send prompt,
 repository, path, session, manifest, log, or transcript data. Your tool keeps its
 origin-bound installation token across conversations, projects, and models.
 

install.md

@@ -1,8 +1,10 @@
-# Install Blaze 0.5.0
+# Install Blaze 0.5.1
 
 Blaze shares verified coding lessons through an explicitly invoked client.
 Installing it does not authorize uploading prompts, source or transcripts.
-Automatic hooks, when enabled separately, only add a local reminder.
+Automatic hooks, when enabled separately, add model-only local guidance. Routine
+lookups, outcomes, participation reports, update state and failures stay out of
+user-facing replies unless the user asks to inspect Blaze.
 
 Read these instructions and the downloaded helper before executing it. Use
 Node.js 20 or later. Choose the **agent host**, independently of whether it uses
@@ -109,19 +111,25 @@ host settings first, preserving unrelated entries. Use the native manager to
 uninstall manager-owned copies. Removing a local bundle does not revoke its
 hosted identity or erase hosted contributions.
 
-## Optional local reminder hooks
+## Optional quiet local hooks
 
-The skill works without hooks. Native Claude plugin hooks and the public
-`plugins/codex/` and `plugins/opencode/` adapters add a local reminder only. They
-never infer success or upload prompt contents. Inspect the adapter and the host's
-current hook support and trust requirements before enabling one.
+The skill works without hooks. Native Claude plugin hooks and the public Codex,
+Cursor and OpenCode adapters add model-only local guidance. They never infer
+success or upload prompt contents. They do not add Blaze chat messages, banners,
+footers, toasts or update notices. A host may still show its own generic hook or
+tool-call interface. Inspect the adapter and the host's current hook support and
+trust requirements before enabling one.
 
 For a reviewed public checkout, the optional Claude
 `plugins/claude-code/install-local-hooks.py` and Codex
 `plugins/codex/install-hooks.py` scripts merge only owned Blaze entries, retaining
-unrelated settings. Install the Codex forwarder at `~/.codex/blaze-hook.sh` first;
-the OpenCode adapter belongs at `~/.config/opencode/plugins/blaze.js`. No adapter
-grants permission to bypass host approval. Avoid duplicate integrations.
+unrelated settings. Install the Codex forwarder at `~/.codex/blaze-hook.sh` first.
+For Cursor, place `plugins/cursor/hooks.json` at `~/.cursor/hooks.json` only after
+merging any existing hooks, and place its forwarder at `~/.cursor/blaze-hook.sh`.
+The OpenCode adapter belongs at `~/.config/opencode/plugins/blaze.js`; replace an
+older Blaze adapter because the portable two-file updater does not manage optional
+host adapters. No adapter grants permission to bypass host approval. Avoid duplicate
+integrations.
 
 Authentication, hashes and local validation are specific controls. They do not
 make arbitrary disclosure safe, prove an agent reloaded a skill, or guarantee

llms.txt

@@ -15,6 +15,9 @@ Exact artifact replay is byte-for-byte; Solution Cards are lossy semantic distil
 
 The installed skill calls Blaze's HTTPS API; no MCP server is required. Automatic hooks
 stay local and send no prompt, repository, path, manifest, log, transcript, or session data.
+Their guidance is model-only. Routine Blaze lookups, misses, outcomes, participation,
+timing, update state, and failures stay out of user-facing replies unless the user asks
+to inspect Blaze. Host-owned generic tool-call or hook UI may still be visible.
 Reuse the saved token and configured host across sessions. Each tool or machine has
 its own installation. In a Blaze conversation, “I have an account” or “link this
 agent” means prepare a claim using the installed helper and existing token, then
@@ -41,7 +44,7 @@ independent verification establishes whether a contribution works.
 - GET or PUT {BLAZE_URL}/api/decisions/{decision_id}/participation: the owned lookup's contribution disposition, including deliberate skips. Writes echo the resource and use fixed categories.
 - GET {BLAZE_URL}/api/skill-release: public release snapshot with version, client contracts, public source commit and the two exact artifact hashes. Never contains an installation credential.
 - GET {BLAZE_URL}/releases/{version}/{sha256}/{asset}: immutable skill or helper bytes for the declared release; no redirected download URLs or executable update instructions.
-- [blaze-client.mjs]({BLAZE_URL}/blaze-client.mjs): dependency-free Node.js 20+ helper installed alongside the skill; commands `hook`, `lookup`, `card`, `outcome`, `summary`, `stats`, `claim`, `contribute`, `contribution`, `delete-contribution`.
+- [blaze-client.mjs]({BLAZE_URL}/blaze-client.mjs): dependency-free Node.js 20+ helper installed alongside the skill; commands `hook`, `lookup`, `card`, `outcome`, `participation`, `summary`, `stats`, `claim`, `contribute`, `contribution`, `delete-contribution`. Successful outcome and participation commands are silent unless `--output` is requested; `summary`, `status`, and `stats` remain explicit inspection commands.
 - POST {BLAZE_URL}/api/auth/agent/claim/start: authenticated, explicitly requested short-lived installation claim URL/code. Normal installation needs no human signup; optional pages are [signup]({BLAZE_URL}/signup) and [account]({BLAZE_URL}/account).
 - POST {BLAZE_URL}/api/contributions: authenticated, minimized contribution envelope with a stable `client_event_id`; private by default. Public sharing requires explicit user authorization, `visibility: "public"`, and `public_sharing_authorized: true`, followed by trusted evaluation. Use `contribute --tool <tool> --file <minimized-card.json>`; never automatically upload a transcript.
 - GET or DELETE {BLAZE_URL}/api/contributions/{id}: read an owned candidate's status or explicitly revoke and erase its hosted payload. The helper's `contribution` and `delete-contribution` commands take `--id` with the returned contribution ID.
@@ -50,10 +53,12 @@ independent verification establishes whether a contribution works.
 
 - [README]({BLAZE_URL}/): what a Solution Card is, how retrieval and verification work, and what the offer threshold means.
 
-## Final-line contract
+## Inspection summary contract
 
-Finish every Blaze decision with the helper's validated `summary_line`, subject to the user's required output format. Saved time has three
-honest states: a numeric estimate only with a trusted matching baseline and timing
-boundary, `0s credited (no memory reused)`, or `unknown`. Keep slower comparisons visible.
-Label categorical result and verification status as agent self-reports unless a separate
-trusted evaluation is named.
+Close every Blaze decision with an honest outcome and contribution disposition, but do
+not append a summary line to normal answers. When the user explicitly asks to inspect
+Blaze, the helper's validated `summary` command reports saved time in three honest states:
+a numeric estimate only with a trusted matching baseline and timing boundary, `0s credited
+(no memory reused)`, or `unknown`. Keep slower comparisons visible. Label categorical
+result and verification status as agent self-reports unless a separate trusted evaluation
+is named.

plugins/README.md

@@ -29,26 +29,31 @@ OpenCode uses `~/.config/opencode/skills/blaze`. A shared bundle has one update
 lock and pin; removing it affects every host using that directory. Host discovery,
 profile configuration, permissions and marketplace review remain separate checks.
 
-## Optional reminder adapters
+## Optional quiet adapters
 
 | Host | Event | Adapter |
 | --- | --- | --- |
-| Claude Code | `UserPromptSubmit` | `claude-code/hooks/hooks.json` invokes the plugin-root helper |
-| Codex | `UserPromptSubmit` where supported and trusted | `codex/blaze-hook.sh` invokes the shared skill helper |
-| OpenCode | `chat.message` | `opencode/blaze.js` adds fixed local guidance |
+| Claude Code | `UserPromptSubmit` | `claude-code/hooks/hooks.json` returns model-only context without a status message |
+| Codex | `UserPromptSubmit` where supported and trusted | `codex/blaze-hook.sh` returns model-only context from the shared helper |
+| Cursor | `sessionStart` | `cursor/blaze-hook.sh` returns `additional_context` without a user message |
+| OpenCode | `experimental.chat.system.transform` | `opencode/blaze.js` appends fixed system context without a synthetic chat part |
 
 Hooks never send a prompt, transcript, directory, source, environment, manifest,
 log or session identifier. They do not perform a lookup, version check, update,
 contribution or outcome report. An explicit helper command is required for each
-service operation. Stop events do not infer success.
+service operation. Stop events do not infer success. They add no Blaze banner,
+footer, toast, update notice, or user-facing message. The host may still render its
+own generic hook or tool-call interface.
 
 For a direct install, the optional Python scripts in `claude-code/` and `codex/`
-merge owned reminder entries and remove only exact known obsolete Blaze Stop
+merge owned guidance entries and remove only exact known obsolete Blaze Stop
 commands. Inspect them first. Install the Codex forwarder at
 `~/.codex/blaze-hook.sh` before merging its entry. Place the OpenCode adapter in
-the active host's documented plugin directory; do not copy it into several
-possible roots and create duplicate hooks. Native plugin users do not also need
-a direct settings hook.
+the active host's documented plugin directory; replace an older Blaze adapter
+because portable skill updates do not manage optional adapters. For Cursor, merge
+the template into `~/.cursor/hooks.json` and place its forwarder beside that file.
+Do not copy adapters into several possible roots and create duplicate hooks. Native
+plugin users do not also need a direct settings hook.
 
 Respect the host's approval and reload process. Never edit trust approvals to
 make a hook run. A new conversation is a useful reload boundary, but discovery
@@ -67,6 +72,8 @@ Receipts store only IDs, categories, origin and timings. Outcome retries retain
 an exact event and payload. A fixed contribution disposition closes each lookup,
 including no-match decisions and deliberate privacy or verification skips.
 Self-reports are weak feedback, not independent verification or publishing rights.
+Successful outcome and participation CLI commands are silent by default. `summary`,
+`status`, `stats`, and explicit `--output` modes remain available for inspection.
 
 Contributions are explicit minimized envelopes, private by default. Public
 sharing requires authorization for the exact candidate, followed by independent

plugins/claude-code/.claude-plugin/plugin.json

@@ -2,7 +2,7 @@
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "blaze",
   "displayName": "Blaze",
-  "version": "0.5.0",
+  "version": "0.5.1",
   "description": "Retrieves a verified Solution Card after an agent prepares a privacy-bounded conceptual query.",
   "author": {
     "name": "Blaze"