POST /api/playground/improve-prompt: validates body via ImprovePromptRequestSchema,
calls /v1/chat/completions internally with the user-chosen model (D8), parses
improved content via parseImprovedContent, returns { improvedSystem?, improvedPrompt?,
tokensIn, tokensOut }. Auth optional (REQUIRE_API_KEY gate). All errors route
through buildErrorBody/sanitizeErrorMessage (Hard Rule #12).
scheduleRecordConsumption() wraps recordConsumption() in setImmediate so it
never adds latency to the client response path. Errors are caught and logged
via pino warn but NEVER propagated to the caller (B29 fail-open contract).
Implements the quota share enforcement gate and consumption recorder:
- enforceQuotaShare(): PRE-request check that returns allow/block/deprioritize
based on fair-share algorithm, saturation signals, and pool allocations.
- recordConsumption(): POST-response tracker that increments per-key counters
for each active plan dimension.
Both functions fail-open per B16/B29: any infra error → allow + warn log.
Adds 5 new settings route handlers for CLIs introduced by plan 14 that
declare configType:"custom" and need automated config file persistence:
forge (~/.forge/config.toml), jcode (~/.jcode/config.json),
deepseek-tui (~/.config/deepseek-tui/config.toml),
smelt (~/.smelt/config.json), pi (~/.pi/config.json).
Also registers the 5 tools in cliRuntime.ts path table so
getCliPrimaryConfigPath() resolves their config paths correctly.
Each handler follows the established pattern: requireCliToolsAuth guard on
every exported method, Zod body validation on POST, buildErrorBody/
sanitizeErrorMessage on all error paths (Hard Rule #12), fs/promises only
(no exec/spawn — Hard Rule #13), saveCliToolLastConfigured on success.
Integration tests: 7 subtests per handler (401 without auth, 200 GET,
400 missing-baseUrl, 400 missing-model, 200 POST writes file, 200 DELETE,
error sanitization + no exec/spawn static audit).
- compliance.actor / compliance.actorPlaceholder added to pt-BR + en
- logs.compressionLogTitle, logs.compressionLogEmpty, logs.tokens copied
from settings to logs namespace in both locales (original keys kept in
settings to avoid breaking other components)
The component was calling useTranslations("settings") which would cause
key misses once the settings namespace is reorganized. Keys used
(loading, compressionLogTitle, compressionLogEmpty, tokens) are now
sourced from the canonical "logs" namespace.
Adds state, fetch param, reset and input+datalist for filtering audit
entries by actor. The actor field is inserted between eventType and
severity in the filter grid. Filter value is sent as ?actor= to the
compliance audit-log API on every change.
Implements POST /api/agent-skills/generate (F4):
- Requires management auth via requireManagementAuth (same pattern as
usage/combo-health-dashboard and other management routes)
- Validates body via GenerateBodySchema (dryRun defaults true, prune
defaults false — safe preview mode)
- Dynamic import of @/lib/agentSkills/generator so the route coexists
before F3 (generator.ts) is merged; returns 503 if module unavailable
- Hard Rule #12: all error paths use buildErrorBody (400/401/403/503/500)
- Hard Rule #7: Zod validates both JSON parse and schema shape
Implements four read-only endpoints for the Agent Skills REST API (F4):
- GET /api/agent-skills — catalog with ?category= and ?area= filters,
returns { skills, count, coverage }
- GET /api/agent-skills/[id] — single skill by canonical ID (404 if absent)
- GET /api/agent-skills/[id]/raw — SKILL.md as text/markdown with
Cache-Control: public, max-age=3600; 502 on GitHub fallback failure
- GET /api/agent-skills/coverage — SkillCoverage (filesystem vs catalog)
All routes: export const dynamic = "force-dynamic" (filesystem reads),
error responses via buildErrorBody (Hard Rule #12), Zod input validation
(Hard Rule #7). coverage/route.ts force-added to git because .gitignore
matches the directory name "coverage/" — this is an API route, not a
report directory.
Adds "list-capabilities" entry to A2A_SKILL_HANDLERS in taskExecution.ts
(dynamic import pattern, consistent with the 5 existing skills) and adds
the 6th skill entry to /.well-known/agent.json with tags [discovery, capabilities]
and example questions for agent discovery.
Implements executeListCapabilities() which calls getCatalog() + computeCoverage()
from the agentSkills catalog (F1/F2) and returns a markdown table covering all
42 skills (22 API + 20 CLI) with ID, name, category, area, endpoints/commands,
and raw SKILL.md URL, matching the §3.7 result contract.
Two additions to src/mitm/manager.ts:
1. writeTargetsJson(targets?) — persists the static ALL_TARGETS
registry to <DATA_DIR>/mitm/targets.json. server.cjs reads this
file at boot and extends its baseline TARGET_HOSTS set so the
full AgentBridge target catalog is intercepted alongside the
historical antigravity hosts.
2. getAllAgentsStatus() — read-only aggregate of every registered
target plus its current installation detection result, used by
the AgentBridge dashboard.
startMitm() now invokes writeTargetsJson() before any DNS/cert work;
write failures are logged but never block startup.
server.cjs now reads <DATA_DIR>/mitm/targets.json at startup and
adds the listed hostnames to TARGET_HOSTS. The antigravity baseline
remains hard-coded so existing installs continue to work even if
targets.json is missing or malformed (loader catches all errors and
returns 0).
All additions are marked with // T-A-F3: comments to make the
forward-port-only changes easy to audit.
Filesystem-only installation probes for antigravity, kiro, copilot,
codex, cursor, zed, claude-code, and open-code. detectAgent(id)
dispatcher returns {installed, path?} without ever spawning a shell
or interpolating runtime paths (Hard Rule #13).
Trae has no detector — its entry in the dispatch table returns
{installed: false} until upstream viability is confirmed.
ALL_TARGETS aggregates the nine MitmTarget descriptors in canonical
order. resolveTarget(hostname) does a case-insensitive exact-match
lookup against each target.hosts list. routeConnection(hostname,
userBypass) returns {kind: bypass|target|passthrough} per plan 11
§4.6 precedence: default+user bypass > known target host > passthrough.
Implements MitmHandlerBase abstract class with shared concerns
(request body capture, secret masking, router forwarding, SSE
piping, Traffic Inspector hooks via dynamic import) plus concrete
handlers for antigravity, kiro, copilot, codex, cursor, zed,
claude-code, open-code, and trae (stub for investigating viability).
Each concrete handler maps request body model field to a configured
target, forwards to OmniRoute router, and pipes back SSE.
Targets antigravity and kiro updated to the new MitmTarget shape
while preserving legacy MITM_PROFILE export aliases.
