fix: skill_runner reads /tmp/skills (matches where install_skills writes) by ericodom · Pull Request #4 · thinkwork-ai/thinkwork

ericodom · 2026-04-12T15:42:32Z

Summary

`skill_runner.py` was pointing at `/app/skills` — a directory the Dockerfile creates empty and never fills. `install_skills.install_skill_from_s3()` downloads every per-request skill to `/tmp/skills`. The read path and the write path never aligned, so every skill silently failed to register on the parent agent.
Found this while validating built-in tools end-to-end. `chat-agent-invoke` logs correctly said "Injected built-in tool 'web-search' (provider=exa)", the AgentCore container received the payload, `install_skill_from_s3` pulled the files down, and then `register_skill_tools_grouped` logged `0 tool-mode tools, 0 agent-mode skills, 0 total` because it was looking in the wrong place. Claude reply: "I don't have a live web search tool available."

Test plan

Repro: observed the `0 tool-mode tools, 0 agent-mode skills` line in /thinkwork/dev/agentcore after sending a chat
After merge + deploy (needs container rebuild): send the same chat, expect Claude to call `web_search` and return actual Austin results
Re-test SerpAPI provider path for the same tool

Depends on the merge of #3.

install_skills.install_skill_from_s3() writes downloaded skills to /tmp/skills/{skill_id}/, but skill_runner.SKILLS_DIR was set to /app/skills — a directory the Dockerfile creates but never populates. Every per-request skill registration therefore silently skipped every skill (parse yields None because skill.yaml doesn't exist under /app). Observed: chat-agent-invoke logs showed 'Injected built-in tool web-search (provider=exa)' and the AgentCore container received the payload, but 'Grouped skill registration: 0 tool-mode tools, 0 agent-mode skills, 0 total' — so Claude never saw web_search, never called it, and replied "I don't have a web search tool available." This fix aligns the read path with the write path. No tenant data lives under /app/ anyway — Lambda /tmp is the only writable location at runtime, and that's already what install_skills uses. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ck (#296) Follow-up to #294 / #295. Marco bootstrap chain worked for the first hop but silently stopped at step 2: job #2 enqueued #3 successfully, but #3 (which hit max_new_pages again) didn't enqueue #4. Root cause: the continuation bucket was computed as `Date.now() + 300s`. When a chained job's own runtime exceeded its bucket length (job ran 112s but bucket is 300s; once runtime plus the offset crosses a bucket boundary, the computed "next" bucket equals the bucket the job itself is running in). The dedupe key collided with the child's own row, `ON CONFLICT DO NOTHING` swallowed the insert, and the chain died without a visible error. Anchor the offset on `job.created_at` instead. Each step in a chain now produces a strictly-monotonic bucket: parent in bucket N ⇒ child enqueued for bucket N+1, child in bucket N+1 ⇒ grandchild for N+2 — regardless of how long any step took to run. Dedupe collisions can only fire against external jobs (e.g., a memory-retain trigger that hit the same bucket), which is the correct behavior. Observed on dev: Marco job chain went 1→2, stopped. Post-fix, re-triggering will exercise the continuation through the full 261 memories until the cursor drains. No test regression (411/419, pre-existing 8 skips). Continuation behavior itself is covered by the Marco integration that spawned this fix.

…tion fix (#318) Closes handoff plan items #3 and #4.1 from plans/2026-04-20-006-handoff-cluster-enrichment-and-followups.md. ## #4.1 extractCityFromAddress dotted-abbreviation fix Spanish/Canadian/Australian addresses like "..., San Miguel de Allende, Gto., Mexico" previously produced candidate "Gto." (Guanajuato) because the region-code walk only recognized `^[A-Z]{2,4}(\s|$)` patterns. The audit showed 32 Marco records producing "Gto." and 22 producing "Q.R." (Quintana Roo) as candidate cities — both now correctly resolve to the preceding city slot. Fix: `isDottedRegionAbbr` recognizes `[A-Z][a-z]{0,2}\.` groups (up to 4 repetitions, ≤ 10 chars) and terminates the walk like the existing US-style region codes do. Record-expander probe on Marco confirms: candidate "San Miguel De Allende" (support=32) replaces "Gto." and the equivalent cities replace "Q.R." ## #3 summary-expander → deterministic linker wiring The audit showed `deriveParentCandidatesFromPageSummaries` produces 91 candidates on Marco (Toronto, Seattle, Honolulu hubs + many more) but the deterministic linker only consumed `deriveParentCandidates(records)` — so those candidates never became links. ### Type extension `DerivedParentCandidate` gets `sourceKind?: "record" | "summary"`: - record: `sourceRecordIds` are memory-record ids (existing path) - summary: `sourceRecordIds` are page ids (new path) Field is optional so existing test fixtures stay green; emitter defaults to "record" when omitted. ### Emitter update `emitDeterministicParentLinks` now builds two leaf indexes: - `leavesByRecord` — keyed on memory-record id (existing) - `leavesById` — keyed on page id, built from `[...scopePages, ...affectedPages]` (new) Summary-kind candidates route through `leavesById`. A new optional `scopePages` arg feeds the index so summary-kind candidates can resolve leaves that weren't touched THIS batch — necessary because summary-based candidates come from a scope-wide scan, not batch-local records. ### Compiler wiring `applyPlan` now calls BOTH expanders and passes merged candidates to the linker, along with `candidatePages` (already fetched for the planner call) as `scopePages`. Bypasses the merge-across-kinds ambiguity by concatenating both lists rather than merging — the emitter handles each candidate according to its kind. ### Precision filter tightening Summary-expander output adds `isLikelyCityToken` filter before pushing into the byCity map: - drops > 4-word fragments ("Prospect Interested In The Full PVL Product Line") - drops < 3-char tokens ("St") - drops street-suffix endings ("Congress Ave", "Queen St") These are the noise categories the 04-20 audit surfaced; real cities like "Buenos Aires" (2 words), "New York" (2 words), and "Montréal" (1 word with accent) all pass through. ## Expected impact On the next Marco recompile: `links_written_deterministic` should increase from the current 14/batch (record-only path) by 30-60 links (Toronto entity leaves → Toronto hub-if-exists-else-fuzzy-match, Seattle / Honolulu / etc.) — validated by unit tests; live-compile numbers land after deploy. Backfill dry-run on Marco: 22 parent links (up from 21), all spot- check precision-correct; wet run: 386 → 387 reference links (+1 net new, most attempts are idempotent re-writes of existing edges). ## Test coverage - 3 new dotted-abbreviation tests (Gto., Q.R., B.C.) - 5 new summary-expander filter tests (word cap, short, street suffix, sourceKind tagging for both expanders) - 5 new deterministic-linker tests for the summary-kind branch (scopePages leaf resolution, empty-pool skip, non-entity filter, back-compat when sourceKind omitted, both-kinds-fire for same parent) - Total: 471 passed / 8 skipped (up from 458). - Typecheck clean. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ks dedup (#319) Captures the 2026-04-20 third-session end state. Supersedes plans/2026-04-20-006-handoff-cluster-enrichment-and-followups.md now that PRs #318 closed out items #3 (summary-expander wiring) and #4.1 (dotted-abbreviation fix). Four remaining items in priority order: 1. Validate #318 on Marco (TOP PRIORITY). PR promised links_written_deterministic jump of +30-60 on the next live compile, but the backfill can't exercise that path. Trigger a manual compile and verify the prediction before building more on top. 2. Unit 6 mention cluster enrichment. Biggest remaining plan item. Brief on the key design decision (mentionClusterEnrichments[] JSON shape — Option A separate rows recommended over Option B inline-promotions). 3. wikiBacklinks dedup. Warm-up PR (<1 hour). listBacklinks misses the dedup pattern that listConnectedPages already has. 4. Trivial grab-bag: wipeWikiScope FK dependency, applier-split debt (good to bundle with #2 since cluster promotion adds more lines to the already-1300-line applyAggregationPlan). Session tally: 6 PRs merged (#309, #311, #312, #316, #317, #318). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Three drift incidents in five days (Apr 17 mig 0008, Apr 18 collision, Apr 21 0018+0019) had the same root cause and the same named-but- unshipped fix. Captures the pattern so the next author sees it before shipping migration #4. The drift-reporter fix shipped in PR #367; this is the learning that explains why the reporter exists and what marker convention every future unindexed migration must follow. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Resolves the P1 tension the handoff flagged: the plan honestly scoped R13 to 'no token via Python-stdio-mediated writes or known-shape CloudWatch patterns' and added T1b (intra-tenant template-author exfil), but the brainstorm still read absolutely. Three edits align the goalposts: 1. R13 rescoped to match the plan, with named residual coverage gaps (os.write at fd level, subprocess env dumps, C-extension writes, multiprocessing workers, adversarial split-writes) tracked as the Stdout-bypass class alongside T1. 2. Success Criterion #4 softened to match — 'within R13's scope.' 3. T1b added as a first-class residual threat between T1 and T2, with v1 mitigations (1-hour TTL, shared-template-author review as compensating control, tenant-as-trust-boundary) and v2 hardening track (per-user ABAC session tags or in-process credential proxy, latter preferred because it addresses T1 and T1b simultaneously).

…ation (Unit 6) (#430) ## createTenant wiring After INSERTing the tenant row, createTenant now invokes the agentcore-admin Lambda's /provision-tenant-sandbox route (plan Unit 5) via a new invokeProvisionTenantSandbox helper. The invoke uses InvocationType: RequestResponse per feedback_avoid_fire_and_forget_lambda_invokes so errors surface inside createTenant — but createTenant catches and logs them so a sandbox outage doesn't turn into a tenant-onboarding outage. The reconciler (Unit 6 follow-up) sweeps rows with null sandbox_interpreter_*_id at its own cadence. invokeProvisionTenantSandbox: - Reads AGENTCORE_ADMIN_LAMBDA_ARN + AGENTCORE_ADMIN_TOKEN env vars; throws SandboxProvisioningConfigError if missing (distinguishable so missing config is a warn, not an error). - Builds the API Gateway v2 envelope the handler is written for. - 45s abort signal matches the handler's own budget. - Translates statusCode 4xx → Error with server-side message; 2xx → structured ProvisionResult. ## updateTenantPolicy (new) Platform-operator-only mutation for sandbox_enabled + compliance_tier policy changes. Separate from updateTenant because the changes are security-boundary shifts and are audited in tenant_policy_events. Gate: caller's email must appear in the THINKWORK_PLATFORM_OPERATOR_EMAILS allowlist (comma-separated env var). This is the swap-out point for formal RBAC when it lands. Transition semantics encoded in a pure computeTransition helper (11 tests): - No-op when nothing changes. - sandbox_enabled true rejected when compliance_tier != standard. - compliance_tier → non-standard coerces sandbox_enabled off, producing a paired audit event so the transition is reproducible from the audit log alone. - tier-first ordering means 'enable sandbox AND set tier to hipaa in one call' deterministically rejects. Writes are wrapped in a db.transaction so the tenants UPDATE and the tenant_policy_events INSERT land atomically. ## GraphQL - Tenant type gains sandboxEnabled / complianceTier / sandboxInterpreter*Id - New UpdateTenantPolicyInput + updateTenantPolicy mutation - pnpm schema:build re-ran against the AppSync subscription schema ## Tests — 16 passing - sandbox-provisioning.test.ts (5) — envelope shape, 200/202/4xx parsing, Lambda FunctionError surfacing - updateTenantPolicy.test.ts (11) — every permutation of the transition decision tree: no-op, toggle true on standard, reject toggle true on regulated/hipaa, toggle false always ok, tier change coerces sandbox, composite tier+sandbox requests ## Deferred to follow-up - Reconciler Lambda (EventBridge scheduled fill + drift passes) — lands when the agentcore-admin Lambda terraform resource lands. Currently handled reactively: sandbox failures on createTenant are logged and the next successful createTenant-invoke retry picks up partial state. - SNS platform-security topic — per handoff P1 #4, dropped from v1 because no named subscriber / SLA exists. Audit row in tenant_policy_events provides detection; notification is v2. - Pre-existing tenants: a one-time operator action flips sandbox_enabled per-tenant during staged rollout (see plan Operational Notes).

…hon skill (#486) Parity pass with packages/skill-catalog/thinkwork-admin/scripts/operations/*.py. The @thinkwork/admin-ops package now exposes every op the Python skill ships, and the admin-ops MCP server registers all of them as MCP tools. Sets up deprecation of the skill: agents using mcp.thinkwork.ai can do everything the skill's Python wrappers did. Client - AdminOpsClient gains a `graphql(query, variables?)` helper that POSTs to /graphql with the same Bearer, throws AdminOpsError on error responses. Mutations + most reads use GraphQL (matches the Python skill's wire); tenants module continues to use REST since those handlers already exist. Ported modules (28 ops): - teams.ts 5 mutations + 2 reads (createTeam, add/remove team agents + users, listTeams, getTeam) - agents.ts 3 mutations + 3 reads (createAgent, setAgentSkills, setAgentCapabilities, listAgents, getAgent, listAllTenantAgents) - templates.ts 5 mutations + 3 reads (createAgentTemplate, createAgentFromTemplate, syncTemplateToAgent, syncTemplateToAllAgents, acceptTemplateUpdate, listTemplates, getTemplate, listLinkedAgentsForTemplate) - users.ts 0 mutations + 3 reads (me, getUser, listTenantMembers) - artifacts.ts 0 mutations + 2 reads (listArtifacts, getArtifact) - _fields.ts shared GraphQL field-selection constants mirroring reads.py MCP tool registration (packages/lambda/admin-ops-mcp.ts) - 25 new tools covering every ported op. Each carries a JSON Schema inputSchema + a non-empty description. Tenant pinning from the authenticated key overrides any caller-supplied tenantId on downstream calls. - The existing tools/list test asserts a curated must-have set rather than an exact-count equality, so future ports don't require test churn. Tests - 6 new tests in packages/admin-ops/src/teams.test.ts covering wire-format correctness (queries contain the right operation names, variables carry through, errors surface as AdminOpsError). - Full monorepo test run: 1270+ tests passing. Not in scope - CLI migration of `thinkwork team/agent/template/user/artifact` subcommands — deferred to a follow-up PR. - Removal of packages/skill-catalog/thinkwork-admin/ — deferred to PR #5 after seed (PR #4) promotes mcp.thinkwork.ai to tenants. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ction preflight (#485) * refactor(sandbox): drop required_connections / OAuth preamble / connection preflight Ends the sandbox's OAuth-into-os.environ path end-to-end. Admin UI already stopped surfacing required_connections in #477; this closes the loop across validator, preflight, dispatcher, container, preamble, the pilot skill, concept doc, and runbook. Why now: the OAuth preamble was a live realization of the T1/T1b/T2 residual-threat classes the concept doc itself warns about. The v2 in-process credential proxy is the planned structural fix; landing that work cleanly requires this path gone. Agents that need OAuth-ed work (Slack, GitHub) call composable-skill connector scripts instead. What survives in the sandbox: - execute_code is a pure-compute primitive - The preamble still runs as executeCode call #1 — now a one-line sitecustomize readiness check that aborts the session if the stdio redactor didn't install (refuses to run user code on an unmitigated image). PREAMBLE_VERSION bumps to 2 - Preflight decision tree shrinks from 5 outcomes to 4; the missing-connection + ConnectionRevoked error classes are gone - Dispatcher payload drops sandbox_secret_paths / sandbox_tenant_id / sandbox_user_id / sandbox_stage; only sandbox_interpreter_id + sandbox_environment survive - packages/api/src/lib/sandbox-secrets.ts deleted outright - Validator rejects required_connections on write (not silently accepts) so operators can't reintroduce it via raw GraphQL - Hydration silently strips the key from legacy DB rows — no migration Docs: - Concept page loses the Preamble section, SandboxMissingConnection and ConnectionRevoked rows, required_connections YAML, and the T1 residual row. T1b dropped since the exfil class no longer exists. The 11-step per-turn lifecycle shrinks to 9 steps - Runbook loses failure modes #3 and #4; architecture-in-one-page simplified accordingly - sandbox-pilot SKILL.md rewritten to pure compute (S3 upload via per-tenant IAM role, no Slack post, no GitHub token) Verification: typecheck clean, 1050 api tests pass, docs site builds 79 pages clean, preamble emission verified via ast.parse + inline regression assertions (no boto3, no SecretString, no os.environ[...], no OAuth env-var names, no token prefixes). Plan: docs/plans/2026-04-23-006-refactor-sandbox-drop-required-connections-plan.md Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * refactor(sandbox): ce-review autofix sweep — tool docstring, e2e fixture, warm-container cleanup Applies ten safe_auto fixes surfaced by the ce:review pass across 9 reviewer personas. Highest-impact fixes: - sandbox_tool.py: rewrote the execute_code Strands tool docstring to describe the pure-compute primitive. Removed the retired OAuth env var claims (GITHUB_ACCESS_TOKEN, SLACK_ACCESS_TOKEN, GCAL_ACCESS_TOKEN) and ConnectionRevoked error — 7 reviewers flagged this; the LLM reads this docstring as live tool guidance - fixtures.ts: dropped required_connections from the e2e fixture's createAgentTemplate call (the validator now rejects it, every integration test would abort in setup). Removed vacuous syntheticTokens / seedConnections / putSecret plumbing that the token-leak assertion's forbiddenValues depended on - sandbox-pilot.e2e.test.ts: kept the structural CloudWatch token-leak check but dropped the synthetic forbiddenValues — no synthetic tokens are injected, so the structural prefix check (ghp_/xoxb-/ ya29./JWT) is the only meaningful guard - invocation_env.py: unconditional pop of all six SANDBOX_* keys at invocation entry. Closes the warm-container carryover window where a pre-deploy invocation with interrupted cleanup could leak stale interpreter id + retired OAuth paths into the next invocation - sandbox-invocation-log.ts: dropped connection_revoked from ALLOWED_EXIT_STATUSES + added a regression test asserting the value is now rejected - sandbox_preamble.py: `if not installed()` → `if installed() is not True` (fail-closed against mocks returning truthy non-True). Dropped `from __future__ import annotations` (no-op after dataclass removal) and f-string with only module-constant interpolation - skill.yaml: rewrote the retired required_connections / OAuth preamble comment block to describe pure-compute + per-tenant IAM S3 access - SKILL.md: replaced non-existent SandboxDisabled error name with accurate description (dispatcher does not register the tool) Verification: - pnpm --filter @thinkwork/api typecheck clean - pnpm --filter @thinkwork/api test — 1051/1051 pass (+1 new guard) - pnpm --filter @thinkwork/docs build — 79 pages clean - Python inline regressions: warm-container stale-key clear; preamble identity check + no OAuth markers; tool docstring clean Residual findings (manual follow-up): - ADV-001: legacy-row hydration — handlers cast agent.sandbox without running through validateTemplateSandbox. Functionally safe (field never read) but plan comment overstates the strip behavior - KT-001: SandboxEnvironmentId hand-rolled duplicate of SandboxEnvironment from database-pg schema - W2: CAPABILITIES.md doesn't mention the sandbox / connector-skill pattern to the agent system prompt Review artifact: .context/compound-engineering/ce-review/20260423-194430-67ea771d/ Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

U4: add Astro redirects: { '/pricing': '/cloud' } in astro.config.mjs and delete pricing.astro. Astro's static build emits /pricing/index.html with <meta http-equiv='refresh'>, <link rel='canonical' href=.../cloud>, and <meta name='robots' content='noindex'> — inbound links (search results, Stripe cancels from older mobile builds, external shares) continue to resolve, and the redirect stub is hidden from search engines so duplicate-content risk is zero. Update every known /pricing reference across the monorepo: - apps/www/src/pages/m/checkout-complete.astro — mobile checkout fallback link "Return to pricing" → "Return to plans" → /cloud. - apps/www/src/env.d.ts — consumer comment refresh. - packages/pricing-config/src/plans.ts — consumer comment refresh. - apps/admin/src/routes/onboarding/welcome.tsx — hardcoded https URL and visible anchor text "Return to pricing" → "Return to plans". - apps/mobile/lib/stripe-checkout.ts — hardcoded cancelUrl to https://thinkwork.ai/cloud. Older installed mobile builds still point at /pricing and will hit the redirect; that's acceptable. - terraform/modules/app/lambda-api/handlers.tf — STRIPE_CHECKOUT_ CANCEL_URL now ends in /cloud so canceled checkouts land directly rather than bouncing through the redirect. Leaves packages/api/src/handlers/stripe-checkout.ts telemetry string 'www-pricing' unchanged per plan Key Technical Decision #4 (analytics continuity).

…te IA (#530) * docs(plan): www Cloud/Services IA refactor plan * feat(www): add /cloud route as copy of pricing page U1: stand up /cloud serving current pricing content. Both routes render identically at this point; subsequent units reframe /cloud, add Services cross-link, and retire /pricing via redirect. * feat(www): rename nav Pricing → Cloud U2: flip the visible nav entry to "Cloud" pointing at /cloud. Single edit in copy.ts propagates to desktop and mobile Header.astro renderings. * feat(www): reframe /cloud as ThinkWork Cloud + add Services cross-link U3: update hero/meta copy from generic "Pricing / Infrastructure you own" to "ThinkWork Cloud / Hosted agent infrastructure, deployed inside your AWS." Add a bulleted clarifier making scope explicit (hosted plans vs separate services vs separate AWS usage vs self-hosted via docs). Add a soft cross-link block between PricingGrid and FinalCTA pointing to /services — prose + inline text link, not a button, to keep the "soft pointer" posture asymmetric with the first-class Cloud Hosting card arriving on /services in U5. PricingGrid and the inline Stripe checkout script are unchanged — the new sections live outside the PricingGrid DOM so [data-plan-cta] / [data-plan-error] selectors continue to match. Also refresh the stale "Do NOT cross-link to /pricing" services-export comment now that the IA split makes the handoff expected in both directions. * refactor: retire /pricing route in favor of /cloud U4: add Astro redirects: { '/pricing': '/cloud' } in astro.config.mjs and delete pricing.astro. Astro's static build emits /pricing/index.html with <meta http-equiv='refresh'>, <link rel='canonical' href=.../cloud>, and <meta name='robots' content='noindex'> — inbound links (search results, Stripe cancels from older mobile builds, external shares) continue to resolve, and the redirect stub is hidden from search engines so duplicate-content risk is zero. Update every known /pricing reference across the monorepo: - apps/www/src/pages/m/checkout-complete.astro — mobile checkout fallback link "Return to pricing" → "Return to plans" → /cloud. - apps/www/src/env.d.ts — consumer comment refresh. - packages/pricing-config/src/plans.ts — consumer comment refresh. - apps/admin/src/routes/onboarding/welcome.tsx — hardcoded https URL and visible anchor text "Return to pricing" → "Return to plans". - apps/mobile/lib/stripe-checkout.ts — hardcoded cancelUrl to https://thinkwork.ai/cloud. Older installed mobile builds still point at /pricing and will hit the redirect; that's acceptable. - terraform/modules/app/lambda-api/handlers.tf — STRIPE_CHECKOUT_ CANCEL_URL now ends in /cloud so canceled checkouts land directly rather than bouncing through the redirect. Leaves packages/api/src/handlers/stripe-checkout.ts telemetry string 'www-pricing' unchanged per plan Key Technical Decision #4 (analytics continuity). * feat(www): consolidate /services into single card grid with Cloud Hosting handoff U5: remove featured/secondary variant split in ServiceCard and services.astro; every card now uses the same visual treatment. Drop the two separate SectionShells (featured grid + 'Additional packages') for a single consolidated section with 5 cards in one 3-column grid (md:grid-cols-2 lg:grid-cols-3): Strategy Sprint, Pilot Launch, Managed Operations, Workflow Expansion, and Cloud Hosting. Governance & Eval + AI Program Advisory are dropped as named cards — Governance substance folds into Managed Operations' includes list; Advisory is cut. Cloud Hosting is a new 5th peer card introducing ctaHref + ctaLabel fields on ServicePackage so it can deep-link visitors to /cloud; other cards leave those fields unset and render without per-card buttons (intake continues via the shared hero and closing CTAs). Preserves id="packages" on the consolidated SectionShell so the hero's #packages anchor link still resolves. Variant discriminator is removed entirely from the ServicePackage type. * refactor(www): correct Cloud positioning + layout polish across /cloud and header Fix the messaging on /cloud after live review: - ThinkWork Cloud is the FULLY-HOSTED product (we operate the platform end-to-end) for teams that don't want to run the Enterprise Agent Harness themselves. The prior framing ("deployed inside your AWS") is the self-managed Enterprise product, not Cloud. - Hero reframed: 'Fully managed AI agents, no infrastructure to run.' with a lede that names the Enterprise Agent Harness as the separate self-managed path. - smallPrint + finePrint block removed from PricingGrid (cards now flow straight into the Services cross-link); the remaining data stays in copy.ts for a potential checkout-page surface. - Cloud Hosting service card on /services reframed: 'Fully managed ThinkWork — no Agent Harness to run' with 'No AWS setup on your side' as its first include bullet. - Shared plan summaries in packages/pricing-config updated so Starter and Enterprise no longer mention deploying inside the customer's AWS (mobile onboarding surfaces these too). Add a /cloud-specific FinalCTA variant (finalCtaCloud in copy.ts) with 'Fully managed' eyebrow, 'Adopt AI. Skip the infrastructure.' headline, and matching points — swapped in via a new 'copy' prop on FinalCTA. Homepage continues to use the default 'Your AWS · Your rules' variant unchanged. Layout polish: - New 'tight' prop on FinalCTA drops its top border + gradient divider and reduces top padding, so /cloud flows cleanly out of the plans section. - RECOMMENDED badge on PricingCard uses opaque #070a0f (matches the page body) instead of bg-brand/10 so the card border no longer shows through the pill. - Services cross-link folded into the same SectionShell as the plans grid with tighter spacing, eliminating an entire empty section between the two. Header.astro now indicates the current route with semibold white + aria-current="page" on desktop and mobile nav; inactive items stay text-slate-400. Trailing slashes and subpaths are handled. * refactor(www): trim /services copy — drop lifecycle section and polish packaging U6: cleanup pass after the packaging consolidation in U5. - Remove the entire 'How it works / Engagement lifecycle' section (services.how in copy.ts + the 4-phase card grid in services.astro). With Strategy Sprint / Pilot Launch / Managed Operations / Workflow Expansion now as packages, the lifecycle narrates the same arc the cards already describe. One telling beats two. - Hero headlineOutcome matches the Services posture from the brief: 'We help teams scope, launch, operate, and expand governed AI workflows.' - Positioning headline de-arced: 'One partner from first workflow to ongoing operations.' Positioning body and meta description also strip the references to Governance & Program Advisory that were dropped in U5. - FAQ loses the 'What happens after launch?' entry — it named the removed packages and restated card content. - Managed Operations card outcome + bestFor trimmed to stop echoing the body. - Packages lede drops the 'Cloud Hosting sits alongside' explanation; the card speaks for itself. With this /services reads as: hero → proof band → positioning → 5-card packages grid → 5-question FAQ → closing CTA.

fix: skill_runner reads /tmp/skills (matches where install_skills writes)