feat(providers): latest Claude models on Anthropic Subscription profile

.ai-run/guides/integration/external-integrations.md

@@ -15,6 +15,7 @@
 | OpenCode | Open-source AI assistant | SSO/API Key | Via CodeMie proxy |
 | MCP Servers | Remote MCP tool servers | OAuth 2.0 (auto) | `codemie-mcp-proxy` |
 | Enterprise SSO | Corporate auth | SAML/OAuth | `SSO_BASE_URL` |
+| Anthropic Subscription | Claude via user's Anthropic account | None (subscription) | `anthropic-subscription` provider |
 
 ---
 
@@ -277,6 +278,43 @@ Claude Code injects `!bash` commands as synthetic `type:'user'` messages. The pr
 
 ---
 
+## Anthropic Subscription Provider
+
+The `anthropic-subscription` provider runs Claude Code against the user's own Anthropic subscription rather than a CodeMie-managed API key. Its model-selection pipeline differs from all other providers.
+
+### Model Is Not Stored in the Profile
+
+Setup omits the `model` field from the profile entirely when none is selected; an empty string is never persisted. The model is chosen per-session inside Claude Code (via `/model`).
+
+| Concern | Behavior |
+|---|---|
+| Profile `model` field | Absent — conditionally omitted at setup, never stored as empty string |
+| Launch banner model line | `CODEMIE_CLI_MODEL` value if set; else `'chosen per session by Claude Code / your Anthropic subscription'` |
+| `models list` command | Prints an informational message; exits 0 — no CodeMie model catalog exists for this provider |
+
+`file:src/providers/plugins/anthropic-subscription/anthropic-subscription.setup-steps.ts:129`
+`file:src/agents/core/launch-model-display.ts:13`
+`file:src/cli/commands/models.ts:82`
+
+### CODEMIE_CLI_MODEL Side-Channel
+
+An explicit `-m/--model` flag writes `CODEMIE_CLI_MODEL` via `applyCliModelEnv()`, clearing any pre-existing value first (stale-shell guard). The `enrichArgs` hook for the `claude` agent reads this var and prepends `--model <value>` to the claude binary argv; a dedup guard prevents double injection when `--model` or `--model=` is already present. Without an explicit `-m`, the var is unset and Claude Code picks the model independently.
+
+| Step | File | Line |
+|---|---|---|
+| Set by AgentCLI | `src/agents/core/AgentCLI.ts` | 210 |
+| Implementation | `src/agents/core/cli-model-env.ts` | 1 |
+| Consumed in enrichArgs | `src/providers/plugins/anthropic-subscription/anthropic-subscription.template.ts` | 103 |
+
+This is the only provider that carries the CLI model via a side-channel env var. All other providers use the `CODEMIE_MODEL → ANTHROPIC_MODEL` pipeline or equivalent.
+
+### Version-Prompt Policy
+
+When installed Claude Code is newer than the CodeMie-verified pin, the upgrade/downgrade prompt defaults to `'continue'` for `anthropic-subscription` (keep the newer binary — model availability depends on it) and `'install'` for all other providers (downgrade to tested pin). The minimum-version hard block is unaffected.
+
+`file:src/agents/core/version-prompt-policy.ts:1`
+`file:src/agents/core/BaseAgentAdapter.ts:472`
+
 ## skills.sh Wrapper (`codemie skills`)
 
 Catalog-agnostic thin wrapper around the upstream `skills` npm CLI. Discovery, ranking, and source classification are out of scope for this CLI.

docs/superpowers/tasks/2026-08-31-subscription-model-access/actual-complexity.json

@@ -0,0 +1,92 @@
+{
+  "task": "Add Anthropic Subscription model access support (EPMCDME-14341): stale-env clearing, per-session model display in launch banner, subscription-aware version-prompt defaults, buildConfig model-omit fix, moonshot guard, and three pure-helper extractions across agent core, CLI commands, and provider plugin layers.",
+  "generated": "2026-08-31T00:00:00Z",
+  "dimensions": {
+    "component_scope": {
+      "score": 4,
+      "label": "L",
+      "affected": "AgentCLI, BaseAgentAdapter, cli-model-env helper, launch-model-display helper, version-prompt-policy helper, models CLI command, setup CLI command, setup-ui utilities, AnthropicSubscriptionTemplate, AnthropicSubscriptionSetupSteps",
+      "layers": "Agent-Tool, Workflow, CLI, Service, Provider-Plugin"
+    },
+    "requirements_clarity": {
+      "score": 2,
+      "label": "S",
+      "status": "Clear",
+      "gaps": "Three code-review fix-ups (stale-env clear, buildConfig omit, moonshot guard) emerged during review rather than upfront; all were low-risk additive changes fully resolved before merge — no unresolved gaps remain"
+    },
+    "technical_risk": {
+      "score": 4,
+      "label": "L",
+      "risk_factors": "Authentication flow touched in AnthropicSubscriptionSetupSteps (Claude browser login orchestration, CodeMie SSO auth); template.ts beforeRun removes ANTHROPIC_AUTH_TOKEN and ANTHROPIC_API_KEY and runs the Claude extension installer; moonshot guard deletes ANTHROPIC_DEFAULT_*_MODEL and CLAUDE_CODE_SUBAGENT_MODEL to prevent cross-provider contamination",
+      "mitigation": "All auth changes are additive conditional branches bounded to the anthropic-subscription provider code path; pure-helper extractions (cli-model-env.ts, launch-model-display.ts, version-prompt-policy.ts) are trivially reversible; existing AgentCLI and BaseAgentAdapter patterns reused throughout with no structural changes to those base classes"
+    },
+    "file_change_estimate": {
+      "score": 6,
+      "label": "XXL",
+      "modified_files": 7,
+      "modified_file_list": [
+        "src/agents/core/AgentCLI.ts",
+        "src/agents/core/BaseAgentAdapter.ts",
+        "src/cli/commands/models.ts",
+        "src/cli/commands/setup.ts",
+        "src/providers/integration/setup-ui.ts",
+        "src/providers/plugins/anthropic-subscription/anthropic-subscription.setup-steps.ts",
+        "src/providers/plugins/anthropic-subscription/anthropic-subscription.template.ts"
+      ],
+      "new_files": 11,
+      "new_file_list": [
+        "src/agents/core/cli-model-env.ts",
+        "src/agents/core/launch-model-display.ts",
+        "src/agents/core/version-prompt-policy.ts",
+        "8 test files (names not provided in changed_files input; ~3 new for the new helpers + ~5 updated for modified sources)"
+      ],
+      "affected_dirs": [
+        "src/agents/core",
+        "src/cli/commands",
+        "src/providers/integration",
+        "src/providers/plugins/anthropic-subscription"
+      ]
+    },
+    "dependencies": {
+      "score": 1,
+      "label": "XS",
+      "new_packages": [],
+      "version_changes": []
+    },
+    "affected_layers": {
+      "score": 4,
+      "label": "L",
+      "layers_changed": ["CLI", "Service", "Agent-Tool", "Provider-Plugin"],
+      "schema_migration": false,
+      "cross_system": false
+    }
+  },
+  "total": 21,
+  "size": "L",
+  "band_range": "21-26",
+  "files_changed": 18,
+  "routing": "brainstorming",
+  "key_reasoning": [
+    {
+      "dimension": "file_change_estimate",
+      "reason": "diffstat N=18 maps to 16+ = XXL: 7 modified source files (AgentCLI.ts, BaseAgentAdapter.ts, models.ts, setup.ts, setup-ui.ts, anthropic-subscription.setup-steps.ts, anthropic-subscription.template.ts) + 3 new pure-helper modules (cli-model-env.ts, launch-model-display.ts, version-prompt-policy.ts) + 8 test files"
+    },
+    {
+      "dimension": "component_scope",
+      "reason": "Bumped M (3) to L (4) by 'affects multiple agents' red flag: applyCliModelEnv() inserted into AgentCLI.ts runs on every agent launch regardless of provider; four distinct component clusters touched: agent core base classes (AgentCLI + BaseAgentAdapter), three new pure-helper modules, two CLI commands, and the full anthropic-subscription plugin (template + setup-steps)"
+    },
+    {
+      "dimension": "technical_risk",
+      "reason": "Bumped M (3) to L (4) by 'affects authentication or authorization' red flag: AnthropicSubscriptionSetupSteps orchestrates Claude browser login and CodeMie SSO auth; template.ts beforeRun hook removes ANTHROPIC_AUTH_TOKEN and ANTHROPIC_API_KEY before the Claude process starts and invokes the extension installer; moonshot guard targets ANTHROPIC_DEFAULT_*_MODEL and CLAUDE_CODE_SUBAGENT_MODEL to block cross-provider contamination"
+    },
+    {
+      "dimension": "affected_layers",
+      "reason": "Four distinct architectural layers: CLI (models.ts + setup.ts commands), Service (setup-ui.ts provider-integration utilities), Agent-Tool/Workflow (AgentCLI + BaseAgentAdapter base classes + three helpers), Provider-Plugin (anthropic-subscription template and setup-steps) — exceeds the M threshold of 2-3 layers; no persistence/migration layer involved"
+    }
+  ],
+  "red_flags_applied": [
+    "Component Scope bumped from M (3) to L (4): applyCliModelEnv() call added to AgentCLI.ts handleRun() runs on every agent launch path — 'affects multiple workflows or agents' red flag.",
+    "Technical Risk bumped from M (3) to L (4): AnthropicSubscriptionSetupSteps orchestrates Claude browser login and CodeMie SSO authentication; template.ts beforeRun removes Anthropic auth env vars from the spawned process environment — 'affects authentication or authorization' red flag."
+  ],
+  "split_recommendation": null
+}

docs/superpowers/tasks/2026-08-31-subscription-model-access/code-review-check.json

@@ -0,0 +1,24 @@
+{
+  "schema": 1,
+  "gate": "code-review.check",
+  "decision": "approve",
+  "confidence": "high",
+  "risk_flags": [],
+  "rationale": "All three prior blocking findings resolved against the fix-up diff (6 files, +91/-6). CR-001: applyCliModelEnv deletes CODEMIE_CLI_MODEL before the conditional set, closing the stale-shell-value leak, with a test covering the shell-preset case. CR-002: buildConfig omits the model field for an empty model and a buildConfig('')->undefined test was added. CR-003: moonshot regression guard added and passing. No new high-risk issue introduced by the fix-up (env helper is a pure delete-then-set; buildConfig spread is clean; remainder is test-only), so no confirmation pass was required. Full unit suite 3969/3969, lint and typecheck clean.",
+  "business_review": [
+    {"criterion": "subscription --model reaches claude; no --model -> default (CS1+CS2)", "status": "pass", "notes": "CR-001 fix removes the stale-value leak; no --model now reliably yields Claude Code default."},
+    {"criterion": "setup not asked for a model; per-session summary (CS6)", "status": "pass", "notes": "CR-002 fix: buildConfig stores no model field."},
+    {"criterion": "D4 tests included", "status": "pass", "notes": "Test-plan items 2 and 5 now implemented."}
+  ],
+  "standards_review": [
+    {"standard": "git-workflow (Conventional Commits + ticket ref)", "status": "pass", "notes": "Fix-up: fix(agents), fix(providers), test(providers), all with Refs: EPMCDME-14341."},
+    {"standard": "code-quality", "status": "pass"},
+    {"standard": "security", "status": "pass"}
+  ],
+  "finding_status": [
+    {"id": "CR-001", "status": "resolved", "notes": "applyCliModelEnv clears before set; tested including shell-preset case."},
+    {"id": "CR-002", "status": "resolved", "notes": "buildConfig omits model field when empty; buildConfig('')->undefined tested."},
+    {"id": "CR-003", "status": "resolved", "notes": "moonshot regression guard added and passing."}
+  ],
+  "findings": []
+}

docs/superpowers/tasks/2026-08-31-subscription-model-access/code-review-final.json

@@ -0,0 +1,54 @@
+{
+  "schema": 1,
+  "gate": "code-review.final",
+  "decision": "request-changes",
+  "confidence": "high",
+  "risk_flags": [],
+  "rationale": "Three lenses ran (blind, edge-case, acceptance); the edge-case lens (repo access) cleared all functional concerns except one confirmed by two lenses. One real correctness bug (CR-001: stale shell CODEMIE_CLI_MODEL leaks into a no---model subscription launch, violating the D3 contract and the 'no --model -> Claude Code default' AC) and two approved-plan test gaps (CR-003 moonshot regression guard; CR-002 buildConfig-empty test + field-omit cleanliness). Dismissed 3 non-issues: --model= dedup form and dual-provider-source (edge-case verified safe for real invocation paths), and an untyped-params false positive from an abbreviated blind-lens diff. Standards clean: 5 Conventional commits with Refs: EPMCDME-14341, code-quality and security pass.",
+  "business_review": [
+    {"criterion": "subscription --model reaches claude; no --model -> default (CS1+CS2)", "status": "partial", "notes": "Set path correct; missing clear-before-set lets a shell-preset value leak (CR-001)."},
+    {"criterion": "model stated to user at launch (CS4)", "status": "pass"},
+    {"criterion": "setup not asked for a model; per-session summary (CS6)", "status": "partial", "notes": "Prompt skipped and summary wired; buildConfig still returns model:'' and test-plan item 5 unmet (CR-002)."},
+    {"criterion": "models list explains source, no unsupported error (CS5)", "status": "pass"},
+    {"criterion": "version: newer->continue default; older note; below-min unchanged; subscription-scoped (CS3/D2)", "status": "pass"},
+    {"criterion": "D1 no entitlement logic; refusal relayed", "status": "pass"},
+    {"criterion": "D4 tests included", "status": "partial", "notes": "Items 2 (moonshot guard) and 5 (buildConfig-empty) missing (CR-002, CR-003)."}
+  ],
+  "standards_review": [
+    {"standard": "git-workflow (Conventional Commits + ticket ref)", "status": "pass", "notes": "5 feat(scope) commits, subjects <100, Refs: EPMCDME-14341 on all."},
+    {"standard": "code-quality", "status": "pass", "notes": "ESM .js imports, explicit return types on new exports, no any, why-only comments."},
+    {"standard": "security", "status": "pass", "notes": "No secrets/injection/unsafe process handling; CODEMIE_CLI_MODEL is a model-id string."}
+  ],
+  "findings": [
+    {
+      "id": "CR-001",
+      "severity": "major",
+      "triage": "patch",
+      "file": "src/agents/core/AgentCLI.ts",
+      "line": 205,
+      "problem": "CODEMIE_CLI_MODEL is set only when options.model is truthy but never cleared first. A value already present in process.env (e.g. shell-exported) survives into anthropic-subscription enrichArgs on a launch with no --model.",
+      "impact": "Violates the D3 env-var contract ('only when the user passed -m/--model this launch') and the AC 'no --model -> Claude Code default': a stale value is silently injected as --model while the banner still shows the per-session phrase.",
+      "recommendation": "Add `delete process.env.CODEMIE_CLI_MODEL;` immediately before the conditional set so each launch starts clean."
+    },
+    {
+      "id": "CR-002",
+      "severity": "major",
+      "triage": "patch",
+      "file": "src/providers/plugins/anthropic-subscription/anthropic-subscription.setup-steps.ts",
+      "line": 126,
+      "problem": "buildConfig still returns model: selectedModel, so the subscription profile persists model:'' rather than omitting the field; spec CS6 says 'stores no model' and test-plan item 5 (buildConfig stores no model) is unimplemented. Functionally inert at runtime (edge-case lens verified) but leaves an empty-string in persisted config.",
+      "impact": "Spec-fidelity gap and an empty-string-in-config value that can surface on a get->update round-trip; the new behavior is untested.",
+      "recommendation": "Omit the field for an empty model: `...(selectedModel ? { model: selectedModel } : {})`; add a test asserting buildConfig(creds, '').model is undefined."
+    },
+    {
+      "id": "CR-003",
+      "severity": "major",
+      "triage": "patch",
+      "file": "src/providers/plugins/moonshot-subscription/__tests__/moonshot-subscription.template.test.ts",
+      "line": 1,
+      "problem": "Spec test-plan item 2 (moonshot-subscription enrichArgs unaffected by CODEMIE_CLI_MODEL) has no test. Scope was verified manually (0 moonshot files changed) but the regression guard the plan committed to is absent.",
+      "impact": "No automated guard that the new CODEMIE_CLI_MODEL mechanism stays anthropic-subscription-only; a future shared-base edit could regress moonshot silently.",
+      "recommendation": "Add a test setting CODEMIE_CLI_MODEL and asserting the moonshot template does not inject --model (unchanged args)."
+    }
+  ]
+}

docs/superpowers/tasks/2026-08-31-subscription-model-access/complexity-assessment.json

@@ -0,0 +1,26 @@
+{
+  "schema": 1,
+  "task": "Fix anthropic-subscription model passthrough, setup messaging, models-list message, and version-prompt default in the @codemieai/code CLI.",
+  "generated": "2026-08-31T00:00:00Z",
+  "dimensions": {
+    "component_scope":      { "score": 4, "label": "L" },
+    "requirements_clarity": { "score": 2, "label": "S" },
+    "technical_risk":       { "score": 3, "label": "M" },
+    "file_change_estimate": { "score": 3, "label": "M" },
+    "dependencies":         { "score": 1, "label": "XS" },
+    "affected_layers":      { "score": 3, "label": "M" }
+  },
+  "total": 16,
+  "size": "M",
+  "routing": "brainstorming",
+  "key_reasoning": [
+    {
+      "dimension": "component_scope",
+      "reason": "Four change sites across three layers: Provider template (anthropic-subscription.template.ts — enrichArgs injection of --model and conditional exportEnvVars blank), Core shared adapter (BaseAgentAdapter.ts — version-prompt default flip from 'install' to 'continue'), and CLI Workflow (models.ts subscription message, setup.ts/setup-ui.ts setup messaging). BaseAgentAdapter is a core shared utility consumed by every agent plugin and provider; the single-line default flip has blast radius across all providers and agents. Red flag applied: touches core shared utility — bumped from M to L."
+    }
+  ],
+  "red_flags_applied": [
+    "Component Scope bumped from M (3) to L (4): BaseAgentAdapter.ts is a core shared utility used by all agent plugins and providers; version-prompt default flip affects every provider's version-check code path."
+  ],
+  "split_recommendation": null
+}

docs/superpowers/tasks/2026-08-31-subscription-model-access/decisions.jsonl

@@ -0,0 +1,4 @@
+{"ts":"2026-08-31T00:00:00Z","gate_id":"spec.approved","mode":"hitl","verdict":{"decision":"approve","rationale":"User approved spec with 6 change sites, CODEMIE_CLI_MODEL passthrough, subscription-scoped version flip, tests included","follow_ups":[],"confidence":"high","source":"hitl"},"escalated":false,"prior_context":{"question":"spec.approved","phase":3,"risk_flags":[],"artifact_refs":["spec.md"]}}
+{"ts":"2026-08-31T00:00:00Z","gate_id":"plan.approved","mode":"hitl","verdict":{"decision":"approve","rationale":"User approved 6-task TDD plan; every impl task test-first with exact anchors and pure-helper extractions","follow_ups":[],"confidence":"high","source":"hitl"},"escalated":false,"prior_context":{"question":"plan.approved","phase":4,"risk_flags":[],"artifact_refs":["plan.md"]}}
+{"ts":"2026-08-31T00:00:00Z","gate_id":"code-review.final","mode":"hitl","verdict":{"decision":"request-changes","rationale":"User chose to fix all 3 findings then re-check","follow_ups":["CR-001","CR-002","CR-003"],"confidence":"high","source":"hitl"},"escalated":false,"prior_context":{"question":"code-review.final","phase":6,"risk_flags":[],"artifact_refs":["code-review.diff","code-review-final.json"]},"prior_orchestrator_verdict":{"decision":"request-changes","confidence":"high","findings":["CR-001","CR-002","CR-003"]}}
+{"ts":"2026-08-31T00:00:00Z","gate_id":"code-review.check","mode":"hitl","verdict":{"decision":"approve","rationale":"All 3 findings resolved; no new issues; suite 3969/3969","follow_ups":[],"confidence":"high","source":"hitl"},"escalated":false,"prior_context":{"question":"code-review.check","phase":6,"risk_flags":[],"artifact_refs":["code-review-check.diff","code-review-check.json"]},"prior_orchestrator_verdict":{"decision":"approve","confidence":"high","finding_status":"CR-001 resolved, CR-002 resolved, CR-003 resolved"}}