… and docs (openclaw#16761)

Add inline file attachment support for sessions_spawn (subagent runtime only):

- Schema: attachments[] (name, content, encoding, mimeType) and attachAs.mountPath hint
- Materialization: files written to .openclaw/attachments/<uuid>/ with manifest.json
- Validation: strict base64 decode, filename checks, size limits, duplicate detection
- Transcript redaction: sanitizeToolCallInputs redacts attachment content from persisted transcripts
- Lifecycle cleanup: safeRemoveAttachmentsDir with symlink-safe path containment check
- Config: tools.sessions_spawn.attachments (enabled, maxFiles, maxFileBytes, maxTotalBytes, retainOnSessionKeep)
- Registry: attachmentsDir/attachmentsRootDir/retainAttachmentsOnKeep on SubagentRunRecord
- ACP rejection: attachments rejected for runtime=acp with clear error message
- Docs: updated tools/index.md, concepts/session-tool.md, configuration-reference.md
- Tests: 85 new/updated tests across 5 test files

Fixes:
- Guard fs.rm in materialization catch block with try/catch (review concern #1)
- Remove unreachable fallback in safeRemoveAttachmentsDir (review concern #7)
- Move attachment cleanup out of retry path to avoid timing issues with announce loop

Co-authored-by: Tyler Yust <TYTYYUST@YAHOO.COM>
Co-authored-by: napetrov <napetrov@users.noreply.github.com>

* feat: add QQ Bot channel extension

* fix(qqbot): add setupWizard to runtime plugin for onboard re-entry

* fix: fix review

* fix: fix review

* chore: sync lockfile and config-docs baseline for qqbot extension

* refactor: 移除图床服务器相关代码

* fix

* docs: 新增 QQ Bot 插件文档并修正链接路径

* refactor: remove credential backup functionality and update setup logic

- Deleted the credential backup module to streamline the codebase.
- Updated the setup surface to handle client secrets more robustly, allowing for configured secret inputs.
- Simplified slash commands by removing unused hot upgrade compatibility checks and related functions.
- Adjusted types to use SecretInput for client secrets in QQBot configuration.
- Modified bundled plugin metadata to allow additional properties in the config schema.

* feat: 添加本地媒体路径解析功能，修正 QQBot 媒体路径处理

* feat: 添加本地媒体路径解析功能，修正 QQBot 媒体路径处理

* feat: remove qqbot-media and qqbot-remind skills, add tests for config and setup

- Deleted the qqbot-media and qqbot-remind skills documentation files.
- Added unit tests for qqbot configuration and setup processes, ensuring proper handling of SecretRef-backed credentials and account configurations.
- Implemented tests for local media path remapping, verifying correct resolution of media file paths.
- Removed obsolete channel and remind tools, streamlining the codebase.

* feat: 更新 QQBot 配置模式，添加音频格式和账户定义

* feat: 添加 QQBot 频道管理和定时提醒技能，更新媒体路径解析功能

* fix

* feat: 添加 /bot-upgrade 指令以查看 QQBot 插件升级指引

* feat: update reminder and qq channel skills

* feat: 更新remind工具投递目标地址格式

* feat: Refactor QQBot payload handling and improve code documentation

- Simplified and clarified the structure of payload interfaces for Cron reminders and media messages.
- Enhanced the parsing function to provide clearer error messages and improved validation.
- Updated platform utility functions for better cross-platform compatibility and clearer documentation.
- Improved text parsing utilities for better readability and consistency in emoji representation.
- Optimized upload cache management with clearer comments and reduced redundancy.
- Integrated QQBot plugin into the bundled channel plugins and updated metadata for installation.

* OK apps/macos/Sources/OpenClaw/HostEnvSecurityPolicy.generated.swift

> openclaw@2026.3.26 check:bundled-channel-config-metadata /Users/yuehuali/code/PR/openclaw
> node --import tsx scripts/generate-bundled-channel-config-metadata.ts --check

[bundled-channel-config-metadata] stale generated output at src/config/bundled-channel-config-metadata.generated.ts
 ELIFECYCLE  Command failed with exit code 1.
 ELIFECYCLE  Command failed with exit code 1.

* feat: 添加 QQBot 渠道配置及相关账户设置

* fix(qqbot): resolve 14 high-priority bugs from PR openclaw#52986 review

DM routing (7 fixes):
- #1: DM slash-command replies use sendDmMessage(guildId) instead of sendC2CMessage(senderId)
- #2: DM qualifiedTarget uses qqbot:dm:${guildId} instead of qqbot:c2c:${senderId}
- #3: sendTextChunks adds DM branch
- #4: sendMarkdownReply adds DM branch for text and Base64 images
- #5: parseAndSendMediaTags maps DM to targetType:dm + guildId
- #6: sendTextToTarget DM branch uses sendDmMessage; MessageTarget adds guildId field
- #7: handleImage/Audio/Video/FilePayload add DM branches

Other high-priority fixes:
- #8: Fix sendC2CVoiceMessage/sendGroupVoiceMessage parameter misalignment
- #9: broadcastMessage uses groupOpenid instead of member_openid for group users
- #10: Unify KnownUser storage - proactive.ts delegates to known-users.ts
- #11: Remove invalid recordKnownUser calls for guild/DM users
- #12: sendGroupMessage uses sendAndNotify to trigger onMessageSent hook
- #13: sendPhoto channel unsupported returns error field
- #14: sendTextAfterMedia adds channel and dm branches

Type fixes:
- DeliverEventContext adds guildId field
- MediaTargetContext.targetType adds dm variant
- sendPlainTextReply imgMediaTarget adds DM branch

* fix(qqbot): resolve 2 blockers + 7 medium-priority bugs from PR openclaw#52986 review

Blocker-1: Remove unused dmPolicy config knob
- dmPolicy was declared in schema/types/plugin.json but never consumed at runtime
- Removed from config-schema.ts, types.ts, and openclaw.plugin.json
- allowFrom remains active (already wired into framework command-auth)

Blocker-2: Gate sensitive slash commands with allowFrom authorization
- SlashCommand interface adds requireAuth?: boolean
- SlashCommandContext adds commandAuthorized: boolean
- /bot-logs set to requireAuth: true (reads local log files)
- matchSlashCommand rejects unauthorized senders for requireAuth commands
- trySlashCommandOrEnqueue computes commandAuthorized from allowFrom config

Medium-priority fixes:
- #15: Strip non-HTTP/non-local markdown image tags to prevent path leakage
- #16: applyQQBotAccountConfig clears clientSecret when setting clientSecretFile and vice versa
- #17: getAdminMarkerFile sanitizes accountId to prevent path traversal
- #18: URGENT_COMMANDS uses exact match instead of startsWith prefix match
- #19: isCronExpression validates each token starts with a cron-valid character
- openclaw#20: --token format validation rejects malformed input without colon separator
- openclaw#21: resolveDefaultQQBotAccountId checks QQBOT_APP_ID environment variable

* test(qqbot): add focused tests for slash command authorization path

- Unauthorized sender rejected for /bot-logs (requireAuth: true)
- Authorized sender allowed for /bot-logs
- Non-requireAuth commands (/bot-ping, /bot-help, /bot-version) work for all senders
- Unknown slash commands return null (passthrough)
- Non-slash messages return null
- Usage query (/bot-logs ?) also gated by auth check

* fix(qqbot): align global TTS fallback with framework config resolution

- Extract isGlobalTTSAvailable to utils/audio-convert.ts, mirroring core
  resolveTtsConfig logic: check auto !== 'off', fall back to legacy
  enabled boolean, default to off when neither is set.
- Add pre-check in reply-dispatcher before calling globalTextToSpeech to
  avoid unnecessary TTS calls and noisy error logs when TTS is not
  configured.
- Remove inline as any casts; use OpenClawConfig type throughout.
- Refactor handleAudioPayload into flat early-return structure with
  unified send path (plugin TTS → global fallback → send).

* fix(qqbot): break ESM circular dependency causing multi-account startup crash

The bundled gateway chunk had a circular static import on the channel
chunk (gateway -> outbound-deliver -> channel, while channel dynamically
imports gateway). When two accounts start concurrently via Promise.all,
the first dynamic import triggers module graph evaluation; the circular
reference causes api exports (including runDiagnostics) to resolve as
undefined before the module finishes evaluating.

Fix: extract chunkText and TEXT_CHUNK_LIMIT from channel.ts into a new
text-utils.ts leaf module. outbound-deliver.ts now imports from
text-utils.ts, breaking the cycle. channel.ts re-exports for backward
compatibility.

* fix(qqbot): serialize gateway module import to prevent multi-account startup race

When multiple accounts start concurrently via Promise.all, each calls
await import('./gateway.js') independently. Due to ESM circular
dependencies in the bundled output, the first import can resolve
transitive exports as undefined before module evaluation completes.

Fix: cache the dynamic import promise in a module-level variable so all
concurrent startAccount calls share the same import, ensuring the
gateway module is fully evaluated before any account uses it.

* refactor(qqbot): remove startup greeting logic

Remove getStartupGreetingPlan and related startup greeting delivery:
- Delete startup-greeting.ts (greeting plan, marker persistence)
- Delete admin-resolver.ts (admin resolution, greeting dispatch)
- Remove startup greeting calls from gateway READY/RESUMED handlers
- Remove isFirstReadyGlobal flag and adminCtx

* fix(qqbot): skip octal escape decoding for Windows local paths

Windows paths like C:\Users\1\file.txt contain backslash-digit sequences
that were incorrectly matched as octal escape sequences and decoded,
corrupting the file path. Detect Windows local paths (drive letter or UNC
prefix) and skip the octal decoding step for them.

* fix bot issue

* feat: 支持 TTS 自动开关并清理配置中的 clientSecretFile

* docs: 添加 QQBot 配置和消息处理的设计说明

* rebase

* fix(qqbot): align slash-command auth with shared command-auth model

Route requireAuth:true slash commands (e.g. /bot-logs) through the
framework's api.registerCommand() so resolveCommandAuthorization()
applies commands.allowFrom.qqbot precedence and qqbot: prefix
normalization before any handler runs.

- slash-commands.ts: registerCommand() now auto-routes by requireAuth
  into two maps (commands / frameworkCommands); getFrameworkCommands()
  exports the auth-required set for framework registration; bot-help
  lists both maps
- index.ts: registerFull() iterates getFrameworkCommands() and calls
  api.registerCommand() for each; handler derives msgType from ctx.from,
  sends file attachments via sendDocument, supports multi-account via
  ctx.accountId
- gateway.ts (inbound): replace raw allowFrom string comparison with
  qqbotPlugin.config.formatAllowFrom() to strip qqbot: prefix and
  uppercase before matching event.senderId
- gateway.ts (pre-dispatch): remove stale auth computation; commandAuthorized
  is true (requireAuth:true commands never reach matchSlashCommand)
- command-auth.test.ts: add regression tests for qqbot: prefix
  normalization in the inbound commandAuthorized computation
- slash-commands.test.ts: update /bot-logs tests to expect null
  (command routed to framework, not in local registry)

* rebase and solve conflict

* fix(qqbot): preserve mixed env setup credentials

---------

Co-authored-by: yuehuali <yuehuali@tencent.com>
Co-authored-by: walli <walli@tencent.com>
Co-authored-by: WideLee <limkuan24@gmail.com>
Co-authored-by: Frank Yang <frank.ekn@gmail.com>

opus-4-7 independent review found three issues missed by earlier
review rounds:

P1 — _persistPathOverride was still module-local
  Despite the Codex thread reply claiming the override moved into the
  shared persistState global, it was still a module-local 'let'. Tests
  passed by accident: in-memory dedupe shared via persistState.entries,
  but disk side effects from re-imported module copies leaked into
  STATE_DIR instead of the per-test temp dir. Moved into
  PersistState.persistPathOverride on the global singleton.

P1 — CWE-59 symlink-follow on tmpPath (Aisle medium #2, unaddressed)
  tmpPath was ${filePath}.${process.pid}.tmp + default 'w' flag. A
  predictable name combined with a truncating O_CREAT would let an
  attacker who can write to the state dir pre-create a symlink at the
  tmp path pointing to e.g. /etc/passwd, and our write would follow it.
  Now uses randomUUID() for the name AND flag 'wx' (O_CREAT | O_EXCL)
  so any existing entry — file or symlink — fails the write safely.

P2 — hydration ignored MAX_ENTRIES (Aisle medium #1)
  The 4 MB file-size cap permits ~250K entries through, blowing past
  the 5K in-memory MAX_ENTRIES on first load. Now collects valid
  candidates first, sorts DESC by ts, and slices to MAX_ENTRIES before
  inserting into the dedupe cache. Bounds memory regardless of file
  size up to MAX_PERSIST_FILE_BYTES.

Tests: 26/26 pass (3 new tests covering each fix).

Architectural note (per opus review): Land standalone. Don't migrate
to plugin-sdk/persistent-dedupe — its hasRecent reads JSON on every
miss, adding ~2-10ms to every Slack message in a thread the bot
hasn't replied to (the dominant path for hasSlackThreadParticipation
on the message-dispatch hot path). Warm-cache + debounced-write is
correct here. Don't block on openclaw#70626.

Five issues raised by Aisle / Codex / Greptile review on PR openclaw#72869,
addressed inline rather than deferred:

1. CWE-59 symlink-following chmod (Aisle high #1)
   ensureModelsFileModeForModelsJson called fs.chmod on a path that
   may be replaced by a symlink. If an attacker can write to the agent
   dir (or OPENCLAW_AGENT_DIR points there), the chmod followed the
   link and changed perms on an arbitrary owned file. Now lstat first
   and refuse to chmod symlinks or non-regular files.

2. Prototype pollution via JSON keys (Aisle medium #3 / CWE-1321)
   stripAuthProfilesVolatileFields() copied untrusted keys into a
   plain {} object. Special keys '__proto__', 'constructor',
   'prototype' could mutate the result's prototype chain. Now uses
   Object.create(null) for the result and explicitly filters those
   three keys (belt-and-suspenders).

3. DoS via unbounded auth-profiles fingerprinting (Aisle medium #4)
   readAuthProfilesStableHash had no size or depth limits.
   - Added MAX_AUTH_PROFILES_BYTES = 8 MiB. Above the cap we hash raw
     bytes instead of running JSON.parse + recursive transform +
     stable-stringify.
   - Added MAX_AUTH_PROFILES_DEPTH = 64 with a depth-cap marker so
     the recursive walk can't stack-overflow on pathologically nested
     input.

4. 'token' incorrectly stripped from fingerprint (Codex P2 / Greptile P2)
   AUTH_PROFILE_VOLATILE_FIELDS included 'token' to keep OAuth session
   token rotation from invalidating the cache. But profiles with
   type: 'token' use the literal 'token' key as a long-lived static
   credential — stripping it would mask real auth-state changes when
   a user rotates a static API token. Removed 'token' from the
   volatile set and documented the boundary inline. OAuth session
   fields ('access', 'refresh') and timing fields stay stripped.

Skipped from this commit (will reply on threads):
- Cache short-circuit on stale on-disk credentials (Codex/Greptile P1):
  separate concern from the security fixes; needs design discussion
  on whether to validate disk-vs-config or remove the short-circuit.
- models.json drift in cache key (Codex P1): same — touches the
  fingerprint shape and overlaps with the targetProvider short-circuit.
- targetProvider short-circuit untested (Greptile P2): test follow-up
  once the short-circuit semantics are settled.
- Aisle medium #5 (raw secrets in fingerprint cache): structurally
  larger refactor; needs to land separately to keep this commit\'s
  blast radius clear.

Lint: 0 errors. TS: clean.

Addresses Codex P1, Greptile P2, and Aisle Med #1/#2/#3 on openclaw#73260:

# Streaming bounded hash (Codex P1 / Aisle Med #2)

The previous oversize-branch in readAuthProfilesStableHash still did
fs.readFile(path) which loaded the entire file into memory \u2014 the
MAX_AUTH_PROFILES_BYTES cap was effectively unenforced.  Same problem
for readModelsJsonContentHash (no cap at all).

New helper safeHashRegularFile():
  - lstat first; reject symlinks and non-regular files
  - bail at lstat-time if size exceeds maxBytes (returns sentinel hash
    so size-change still invalidates the cache)
  - open with O_NOFOLLOW where supported (closes lstat\u2192open TOCTOU)
  - fstat after open to verify it's still a regular file
  - createReadStream(fd) with bounded highWaterMark; destroy if accumulated
    bytes exceed maxBytes (handles the case where attacker grows the
    file between lstat and stream completion)

Both readers now route through safeHashRegularFile.  Models.json gets
a 1 MiB cap (MAX_MODELS_JSON_BYTES); auth-profiles keeps its 8 MiB cap.

# O_NOFOLLOW chmod (Aisle Med #1, CWE-367)

The previous lstat-then-chmod sequence was racy: an attacker who
could rename/replace ${agentDir}/models.json between the two syscalls
could have chmod() follow a swapped-in symlink to an arbitrary file.

Replaced with fs.open(path, O_RDONLY | O_NOFOLLOW) \u2192 fchmod via the
file handle.  The open itself refuses symlinks atomically; the
fchmod operates on the validated fd, eliminating the race.  When
O_NOFOLLOW is unavailable (rare but possible), falls back to plain
O_RDONLY \u2014 the prior lstat protection was already best-effort.

# Symlink-safe reads (Aisle Med #3, CWE-59)

Previous fs.readFile in the hash readers followed symlinks.  With
the safeHashRegularFile refactor above, both auth-profiles.json and
models.json now reject symlinks before reading.

# JSDoc fix (Greptile P2)

readModelsJsonContentHash's prior comment said "forces a re-plan"
when the file is absent, but the call site does null === null
comparison which is a cache hit, not a re-plan (the steady-state
skip case for plan.action === 'skip').  Updated the comment to
explain the actual semantics: stable absence is a stable result;
drift is a re-plan.

# Tests

14/14 existing tests pass.  No new test cases needed \u2014 the existing
file-mode and fingerprint-cache tests cover both the symlink-rejection
paths (via lstat shortcut) and the modelsJsonHash second-factor.
Lint: 0 errors.

…tegration

Addresses Codex P2, Greptile P1+P2x2, and Aisle High #1 + Med #2/#3/#4
on PR openclaw#73261:

# Greptile P1 / Aisle High #1: asymmetric baseUrl (CWE-918, SSRF)

The previous guard:
  if (typeof configuredProvider.baseUrl === 'string' &&
      configuredProvider.baseUrl !== diskProvider.baseUrl)
short-circuited the check entirely when config omitted baseUrl
(common for providers with compiled-in defaults).  An attacker who
could tamper with on-disk models.json could set baseUrl to an exfil
endpoint and the short-circuit would accept it silently \u2014 exactly
the SSRF/credential-exfil vector this PR was meant to close.

Replaced with symmetric stableEqual() so config-undefined vs
disk-string is a mismatch and falls through to full planning, which
re-applies provider/plugin defaults and rewrites the file.

# Codex P2: env-ref API key comparison

resolveConfiguredApiKeyForCompare resolved env refs to the env-var
VALUE (env[ref.id]).  But planOpenClawModelsJson persists env-source
api keys to models.json as the env-var NAME (e.g. 'OPENAI_API_KEY'),
not the value \u2014 that's what resolveApiKeyFromCredential returns for
env source.  Comparing value-vs-name always mismatched, so the
short-circuit never fired for the most common config form
('apiKey: "${env.OPENAI_API_KEY}"').

Now the helper returns the env-var NAME for env-source refs, while
still verifying the env is currently populated (so we don't
short-circuit on a misconfigured environment).  Plaintext values
still compare directly.

# Greptile P2 (perf): readyCache integration

The short-circuit returned before the readyCache check, so warm
callers paid the disk read + JSON.parse cost on every call.
Reordered:
  1. compute fingerprint (cheap)
  2. check readyCache \u2014 warm hit returns immediately, no disk I/O
  3. if cold, attempt short-circuit on disk
  4. if short-circuit succeeds, populate readyCache so subsequent
     calls take the warm path

Net effect: warm callers now skip disk entirely; cold callers with
intact disk state still get the short-circuit benefit; full plan
fires only on real drift.

# Aisle Med #2 (CWE-59): symlink-following short-circuit reads

readExistingProviderMatchesConfig used fs.readFile on a possibly-
symlinked target.  Now lstat-checks first and refuses symlinks /
non-regular files.

# Aisle Med #3 (CWE-1321): prototype-chain key access

explicitProviders[targetProvider] could fall through to a prototype
key like '__proto__' / 'constructor' / 'prototype'.  Now uses
Object.hasOwn at both lookup sites (caller AND
readExistingProviderMatchesConfig) to refuse inherited keys.  Also
explicit string check rejects the three known dangerous keys.

# Aisle Med #4 (CWE-400): unbounded models.json read

Added MAX_MODELS_JSON_SHORT_CIRCUIT_BYTES = 1 MiB cap on the disk
read in the short-circuit path.  Files above the cap fall through
to full planning rather than being parsed.

# Tests

Updated hit-after-warm-fingerprint test (Greptile P2) to use
fs.readFile spy and assert no disk read occurs on the warm path.
Added short-circuit-populates-cache test that drops the in-memory
cache between runs, fires the short-circuit, then verifies the
third call takes the warm path with no disk read.

14/14 tests pass.  Lint: 0 errors.

…w feedback)

Codex P2 / Aisle medium #1 on PR openclaw#73260 flagged that `safeHashRegularFile`
returned a deterministic `oversize:${lst.size}` sentinel when a file
exceeded the cap, so an attacker could swap the *contents* of an oversized
`auth-profiles.json` or `models.json` without changing its byte length and
still hit the readyCache via the size-only sentinel comparison (CWE-345).

Switch to fail-closed:

- `safeHashRegularFile` now returns `null` for oversize at lstat time and
  for grow-past-cap mid-read (the latter was already caught via the
  destroy-with-error path; the JSDoc/comment now states that explicitly).
- The return type narrows from `{ hash; raw: Buffer | null } | null` to
  `{ hash; raw: Buffer } | null` — the `raw === null` branch in
  `readAuthProfilesStableHash` is now unreachable and removed.
- JSDoc on the helper and `readAuthProfilesStableHash` now describes the
  fail-closed semantics and links them back to the threat model.

Also:

- Add a regression test that warms the cache with a small, hashable
  `auth-profiles.json`, then grows it past `MAX_AUTH_PROFILES_BYTES`
  (8 MiB) and asserts the implicit-provider-discovery pipeline re-runs
  (i.e. the fingerprint changed instead of stably hitting an
  `oversize:<size>` sentinel).
- Add the missing `Unreleased` CHANGELOG entry covering the
  content-hashed fingerprint, post-write models.json drift hash, the
  oversize fail-closed change, and the existing `O_NOFOLLOW`/streaming/
  prototype-pollution hardening (Codex P3 on the same review).
- Backfill the new `resolveProviderEnvAuthEvidence`/
  `listProviderEnvAuthLookupKeys`/`resolveProviderEnvAuthLookupKeys`
  exports in the fingerprint test's `model-auth-env-vars.js` mock so the
  suite runs at all after the recent `origin/main` merge added those
  call sites in `model-auth-env.ts` / provider secret-helpers.

Validation (from
~/projects/worktrees/openclaw/perf-models-config-cache-fingerprint):

  pnpm vitest run src/agents/models-config.fingerprint-cache.test.ts
    Test Files  2 passed (2)
         Tests  12 passed (12)
  pnpm tsgo:core            # 0 errors
  pnpm tsgo:core:test       # 0 errors
  pnpm oxlint src/agents/models-config.ts \
              src/agents/models-config.fingerprint-cache.test.ts
    Found 0 warnings and 0 errors.
  git diff --check          # clean

Beads: openclaw-9i0