job-forge 2.14.11 → 2.14.12

package/.claude/agents/general-free.md

@@ -1,10 +1,10 @@
 ---
 name: general-free
-description: Procedural worker on free-tier model. Use for form filling via Geometra, tracker updates, TSV merges, scan dedup, OTP retrieval, and other mechanical/scripted tasks where quality-sensitive text generation is NOT required.
+description: Procedural worker on the low-cost DeepSeek V4 Flash OpenCode route. Use for form filling via Geometra, tracker updates, TSV merges, scan dedup, OTP retrieval, and other mechanical/scripted tasks where quality-sensitive text generation is NOT required.
 model: claude-haiku-4-5
 ---
 
-You are the @general-free subagent. You run on a free-tier model, which means the orchestrator has delegated this task to you **specifically because the work is procedural**: deterministic steps, scripted outputs, no nuanced writing required.
+You are the @general-free subagent. You run on JobForge's low-cost procedural model, which means the orchestrator has delegated this task to you **specifically because the work is procedural**: deterministic steps, scripted outputs, no nuanced writing required.
 
 ## Run This Pre-Flight First Every Time
 

package/.claude/agents/general-paid.md

@@ -1,15 +1,15 @@
 ---
 name: general-paid
-description: Quality-sensitive worker on the strongest free-tier OpenCode model by default. Use for offer evaluation narratives (Blocks A-F), cover letter generation, "Why X?" form answers, interview STAR stories, and other tasks where writing quality and judgment matter.
+description: Quality-sensitive worker on the low-cost DeepSeek V4 Flash OpenCode route by default. Use for offer evaluation narratives (Blocks A-F), cover letter generation, "Why X?" form answers, interview STAR stories, and other tasks where writing quality and judgment matter.
 model: claude-opus-4-7
 ---
 
 You are the @general-paid subagent. The orchestrator delegated this task to you because it requires quality writing or judgment — the kind of work `@general-free` isn't well-suited for.
 
-On OpenCode, this agent now defaults to a free OpenRouter model. On other
-harnesses, the same role may still resolve to a premium model. Your job is
-still the same: produce the best final writing you can from the context you
-were given.
+On OpenCode, this agent defaults to DeepSeek V4 Flash so application work
+does not fall back into overloaded free OpenRouter pools. On other harnesses,
+the same role may still resolve to a premium model. Your job is still the
+same: produce the best final writing you can from the context you were given.
 
 ## Do These Tasks
 

package/.claude/agents/glm-minimal.md

@@ -1,6 +1,6 @@
 ---
 name: glm-minimal
-description: Narrow-scope extractor on free-tier model. Use for single-purpose tasks where the orchestrator passes the exact input and expects a small, structured output — e.g., "extract these 8 fields from this JD text" or "parse this form schema into a label→type map". NOT for multi-step workflows.
+description: Narrow-scope extractor on the low-cost DeepSeek V4 Flash OpenCode route. Use for single-purpose tasks where the orchestrator passes the exact input and expects a small, structured output — e.g., "extract these 8 fields from this JD text" or "parse this form schema into a label→type map". NOT for multi-step workflows.
 model: claude-haiku-4-5
 ---
 

package/.cursor/rules/agent-general-free.mdc

@@ -1,9 +1,9 @@
 ---
-description: Procedural worker on free-tier model. Use for form filling via Geometra, tracker updates, TSV merges, scan dedup, OTP retrieval, and other mechanical/scripted tasks where quality-sensitive text generation is NOT required.
+description: Procedural worker on the low-cost DeepSeek V4 Flash OpenCode route. Use for form filling via Geometra, tracker updates, TSV merges, scan dedup, OTP retrieval, and other mechanical/scripted tasks where quality-sensitive text generation is NOT required.
 alwaysApply: false
 ---
 
-You are the @general-free subagent. You run on a free-tier model, which means the orchestrator has delegated this task to you **specifically because the work is procedural**: deterministic steps, scripted outputs, no nuanced writing required.
+You are the @general-free subagent. You run on JobForge's low-cost procedural model, which means the orchestrator has delegated this task to you **specifically because the work is procedural**: deterministic steps, scripted outputs, no nuanced writing required.
 
 ## Run This Pre-Flight First Every Time
 

package/.cursor/rules/agent-general-paid.mdc

@@ -1,14 +1,14 @@
 ---
-description: Quality-sensitive worker on the strongest free-tier OpenCode model by default. Use for offer evaluation narratives (Blocks A-F), cover letter generation, "Why X?" form answers, interview STAR stories, and other tasks where writing quality and judgment matter.
+description: Quality-sensitive worker on the low-cost DeepSeek V4 Flash OpenCode route by default. Use for offer evaluation narratives (Blocks A-F), cover letter generation, "Why X?" form answers, interview STAR stories, and other tasks where writing quality and judgment matter.
 alwaysApply: false
 ---
 
 You are the @general-paid subagent. The orchestrator delegated this task to you because it requires quality writing or judgment — the kind of work `@general-free` isn't well-suited for.
 
-On OpenCode, this agent now defaults to a free OpenRouter model. On other
-harnesses, the same role may still resolve to a premium model. Your job is
-still the same: produce the best final writing you can from the context you
-were given.
+On OpenCode, this agent defaults to DeepSeek V4 Flash so application work
+does not fall back into overloaded free OpenRouter pools. On other harnesses,
+the same role may still resolve to a premium model. Your job is still the
+same: produce the best final writing you can from the context you were given.
 
 ## Do These Tasks
 

package/.cursor/rules/agent-glm-minimal.mdc

@@ -1,5 +1,5 @@
 ---
-description: Narrow-scope extractor on free-tier model. Use for single-purpose tasks where the orchestrator passes the exact input and expects a small, structured output — e.g., "extract these 8 fields from this JD text" or "parse this form schema into a label→type map". NOT for multi-step workflows.
+description: Narrow-scope extractor on the low-cost DeepSeek V4 Flash OpenCode route. Use for single-purpose tasks where the orchestrator passes the exact input and expects a small, structured output — e.g., "extract these 8 fields from this JD text" or "parse this form schema into a label→type map". NOT for multi-step workflows.
 alwaysApply: false
 ---
 

package/.cursor/rules/main.mdc

@@ -10,7 +10,7 @@ AI-powered job search pipeline: scans portals, evaluates offers, generates CVs v
 ## Hard limits
 
 - [H1] Max 2 parallel `task` dispatches per message. For N jobs, run `ceil(N/2)` sequential rounds of 2. A round is not complete until both subagents return a final outcome (`APPLIED`, `APPLY FAILED`, `SKIP`, `Discarded`, or a written TSV path). A `task` tool result that only gives a session id / title is a launch acknowledgement, not completion. Applies in all modes, for all user phrasings ("urgent", "apply to 10 jobs now").
-  why: higher parallelism blows through free-tier rate limits; each subagent requires post-cleanup and racing more than 2 reliably loses at least one result. On 2026-04-25 the orchestrator launched round 2 while round 1 had only returned task ids, leaving four application subagents in flight and losing two provider-fallback recoveries
+  why: each subagent requires post-cleanup and racing more than 2 reliably loses at least one result. On 2026-04-25 the orchestrator launched round 2 while round 1 had only returned task ids, leaving four application subagents in flight and losing two provider recoveries
 
 - [H2] Max 1 application per company+role. Before every `apply` dispatch, grep all four sources for the URL and for `company+role`: `data/pipeline.md`, all `data/applications/*.md` day files, `batch/tracker-additions/*.tsv`, `batch/tracker-additions/merged/*.tsv`. If any source shows APPLIED / Applied, skip the dispatch.
   why: 2026-04 same-day batch collision — when two batches target the same role, `npx job-forge merge` updates the existing day-file row rather than appending, so grepping day files alone misses earlier-batch applies; merged/*.tsv is the only place the breadcrumb remains
@@ -42,7 +42,7 @@ AI-powered job search pipeline: scans portals, evaluates offers, generates CVs v
   why: iso-trace showed 0.25% Agent calls across 5174 turns under a prior over-broad "delegate before 2nd tool call" rule — the rule was ignored in practice; narrowing matches the original cache-bust incident
 
 - [D2] Route subagent work by cost tier. `@general-free`: procedural — form-fill, TSV merge, verify, OTP retrieval, portal scan metadata extraction, one-shot structured-field transforms. `@general-paid`: quality-sensitive — offer evaluation narrative Blocks A-F, cover letters, "Why X?" answers, STAR interview stories, LinkedIn outreach. `@glm-minimal`: narrow ≤5K-input one-shot extract/classify jobs that do not need context.
-  why: GLM 5.1 doesn't discount cache reads so procedural work there costs ~10×; free-tier models handle procedural work fine empirically (`opencode/big-pickle` processed 1000+ messages at $0)
+  why: OpenCode routes all JobForge tiers through DeepSeek V4 Flash by default now; recent traces showed free OpenRouter fallbacks freezing or hitting provider balance errors during applications
 
 - [D3] Read the active mode file before dispatch. Mode files own score gates, provider fallback, portal runbooks, and output shape.
   why: mode-specific rules change faster than global orchestration rules; keeping them out of the shared prefix preserves cache efficiency and prevents stale branches

package/.opencode/agents/general-free.md

@@ -1,7 +1,7 @@
 ---
-description: Procedural worker on free-tier model. Use for form filling via Geometra, tracker updates, TSV merges, scan dedup, OTP retrieval, and other mechanical/scripted tasks where quality-sensitive text generation is NOT required.
+description: Procedural worker on the low-cost DeepSeek V4 Flash OpenCode route. Use for form filling via Geometra, tracker updates, TSV merges, scan dedup, OTP retrieval, and other mechanical/scripted tasks where quality-sensitive text generation is NOT required.
 mode: subagent
-model: opencode/big-pickle
+model: opencode-go/deepseek-v4-flash
 tools:
   geometra_connect: true
   geometra_page_model: true
@@ -17,14 +17,9 @@ tools:
   task: false
 temperature: 0.1
 reasoningEffort: minimal
-fallback_models:
-  - openrouter/z-ai/glm-4.5-air:free
-  - openrouter/openai/gpt-oss-20b:free
-  - openrouter/nvidia/nemotron-3-nano-30b-a3b:free
-  - openrouter/qwen/qwen3-coder:free
 ---
 
-You are the @general-free subagent. You run on a free-tier model, which means the orchestrator has delegated this task to you **specifically because the work is procedural**: deterministic steps, scripted outputs, no nuanced writing required.
+You are the @general-free subagent. You run on JobForge's low-cost procedural model, which means the orchestrator has delegated this task to you **specifically because the work is procedural**: deterministic steps, scripted outputs, no nuanced writing required.
 
 ## Run This Pre-Flight First Every Time
 

package/.opencode/agents/general-paid.md

@@ -1,7 +1,7 @@
 ---
-description: Quality-sensitive worker on the strongest free-tier OpenCode model by default. Use for offer evaluation narratives (Blocks A-F), cover letter generation, "Why X?" form answers, interview STAR stories, and other tasks where writing quality and judgment matter.
+description: Quality-sensitive worker on the low-cost DeepSeek V4 Flash OpenCode route by default. Use for offer evaluation narratives (Blocks A-F), cover letter generation, "Why X?" form answers, interview STAR stories, and other tasks where writing quality and judgment matter.
 mode: subagent
-model: openrouter/qwen/qwen3-next-80b-a3b-instruct:free
+model: opencode-go/deepseek-v4-flash
 tools:
   geometra_connect: true
   geometra_page_model: true
@@ -17,21 +17,14 @@ tools:
   task: false
 temperature: 0.3
 reasoningEffort: medium
-fallback_models:
-  - openrouter/openai/gpt-oss-120b:free
-  - openrouter/nvidia/nemotron-3-super-120b-a12b:free
-  - openrouter/z-ai/glm-4.5-air:free
-  - openrouter/qwen/qwen3-coder:free
-  - openrouter/google/gemma-4-31b-it:free
-  - openrouter/meta-llama/llama-3.3-70b-instruct:free
 ---
 
 You are the @general-paid subagent. The orchestrator delegated this task to you because it requires quality writing or judgment — the kind of work `@general-free` isn't well-suited for.
 
-On OpenCode, this agent now defaults to a free OpenRouter model. On other
-harnesses, the same role may still resolve to a premium model. Your job is
-still the same: produce the best final writing you can from the context you
-were given.
+On OpenCode, this agent defaults to DeepSeek V4 Flash so application work
+does not fall back into overloaded free OpenRouter pools. On other harnesses,
+the same role may still resolve to a premium model. Your job is still the
+same: produce the best final writing you can from the context you were given.
 
 ## Do These Tasks
 

package/.opencode/agents/glm-minimal.md

@@ -1,7 +1,7 @@
 ---
-description: Narrow-scope extractor on free-tier model. Use for single-purpose tasks where the orchestrator passes the exact input and expects a small, structured output — e.g., "extract these 8 fields from this JD text" or "parse this form schema into a label→type map". NOT for multi-step workflows.
+description: Narrow-scope extractor on the low-cost DeepSeek V4 Flash OpenCode route. Use for single-purpose tasks where the orchestrator passes the exact input and expects a small, structured output — e.g., "extract these 8 fields from this JD text" or "parse this form schema into a label→type map". NOT for multi-step workflows.
 mode: subagent
-model: openrouter/openai/gpt-oss-20b:free
+model: opencode-go/deepseek-v4-flash
 tools:
   geometra_*: false
   gmail_*: false
@@ -13,11 +13,6 @@ tools:
   task: false
 temperature: 0
 reasoningEffort: none
-fallback_models:
-  - openrouter/google/gemma-4-26b-a4b-it:free
-  - openrouter/nvidia/nemotron-nano-9b-v2:free
-  - openrouter/google/gemma-4-31b-it:free
-  - openrouter/z-ai/glm-4.5-air:free
 ---
 
 You are the @glm-minimal subagent. You handle narrow, one-shot extractions where the orchestrator has pre-digested the context and just needs you to do a specific transform.

package/AGENTS.md

@@ -5,7 +5,7 @@ AI-powered job search pipeline: scans portals, evaluates offers, generates CVs v
 ## Hard limits
 
 - [H1] Max 2 parallel `task` dispatches per message. For N jobs, run `ceil(N/2)` sequential rounds of 2. A round is not complete until both subagents return a final outcome (`APPLIED`, `APPLY FAILED`, `SKIP`, `Discarded`, or a written TSV path). A `task` tool result that only gives a session id / title is a launch acknowledgement, not completion. Applies in all modes, for all user phrasings ("urgent", "apply to 10 jobs now").
-  why: higher parallelism blows through free-tier rate limits; each subagent requires post-cleanup and racing more than 2 reliably loses at least one result. On 2026-04-25 the orchestrator launched round 2 while round 1 had only returned task ids, leaving four application subagents in flight and losing two provider-fallback recoveries
+  why: each subagent requires post-cleanup and racing more than 2 reliably loses at least one result. On 2026-04-25 the orchestrator launched round 2 while round 1 had only returned task ids, leaving four application subagents in flight and losing two provider recoveries
 
 - [H2] Max 1 application per company+role. Before every `apply` dispatch, grep all four sources for the URL and for `company+role`: `data/pipeline.md`, all `data/applications/*.md` day files, `batch/tracker-additions/*.tsv`, `batch/tracker-additions/merged/*.tsv`. If any source shows APPLIED / Applied, skip the dispatch.
   why: 2026-04 same-day batch collision — when two batches target the same role, `npx job-forge merge` updates the existing day-file row rather than appending, so grepping day files alone misses earlier-batch applies; merged/*.tsv is the only place the breadcrumb remains
@@ -37,7 +37,7 @@ AI-powered job search pipeline: scans portals, evaluates offers, generates CVs v
   why: iso-trace showed 0.25% Agent calls across 5174 turns under a prior over-broad "delegate before 2nd tool call" rule — the rule was ignored in practice; narrowing matches the original cache-bust incident
 
 - [D2] Route subagent work by cost tier. `@general-free`: procedural — form-fill, TSV merge, verify, OTP retrieval, portal scan metadata extraction, one-shot structured-field transforms. `@general-paid`: quality-sensitive — offer evaluation narrative Blocks A-F, cover letters, "Why X?" answers, STAR interview stories, LinkedIn outreach. `@glm-minimal`: narrow ≤5K-input one-shot extract/classify jobs that do not need context.
-  why: GLM 5.1 doesn't discount cache reads so procedural work there costs ~10×; free-tier models handle procedural work fine empirically (`opencode/big-pickle` processed 1000+ messages at $0)
+  why: OpenCode routes all JobForge tiers through DeepSeek V4 Flash by default now; recent traces showed free OpenRouter fallbacks freezing or hitting provider balance errors during applications
 
 - [D3] Read the active mode file before dispatch. Mode files own score gates, provider fallback, portal runbooks, and output shape.
   why: mode-specific rules change faster than global orchestration rules; keeping them out of the shared prefix preserves cache efficiency and prevents stale branches

package/CLAUDE.md

@@ -5,7 +5,7 @@ AI-powered job search pipeline: scans portals, evaluates offers, generates CVs v
 ## Hard limits
 
 - [H1] Max 2 parallel `task` dispatches per message. For N jobs, run `ceil(N/2)` sequential rounds of 2. A round is not complete until both subagents return a final outcome (`APPLIED`, `APPLY FAILED`, `SKIP`, `Discarded`, or a written TSV path). A `task` tool result that only gives a session id / title is a launch acknowledgement, not completion. Applies in all modes, for all user phrasings ("urgent", "apply to 10 jobs now").
-  why: higher parallelism blows through free-tier rate limits; each subagent requires post-cleanup and racing more than 2 reliably loses at least one result. On 2026-04-25 the orchestrator launched round 2 while round 1 had only returned task ids, leaving four application subagents in flight and losing two provider-fallback recoveries
+  why: each subagent requires post-cleanup and racing more than 2 reliably loses at least one result. On 2026-04-25 the orchestrator launched round 2 while round 1 had only returned task ids, leaving four application subagents in flight and losing two provider recoveries
 
 - [H2] Max 1 application per company+role. Before every `apply` dispatch, grep all four sources for the URL and for `company+role`: `data/pipeline.md`, all `data/applications/*.md` day files, `batch/tracker-additions/*.tsv`, `batch/tracker-additions/merged/*.tsv`. If any source shows APPLIED / Applied, skip the dispatch.
   why: 2026-04 same-day batch collision — when two batches target the same role, `npx job-forge merge` updates the existing day-file row rather than appending, so grepping day files alone misses earlier-batch applies; merged/*.tsv is the only place the breadcrumb remains
@@ -37,7 +37,7 @@ AI-powered job search pipeline: scans portals, evaluates offers, generates CVs v
   why: iso-trace showed 0.25% Agent calls across 5174 turns under a prior over-broad "delegate before 2nd tool call" rule — the rule was ignored in practice; narrowing matches the original cache-bust incident
 
 - [D2] Route subagent work by cost tier. `@general-free`: procedural — form-fill, TSV merge, verify, OTP retrieval, portal scan metadata extraction, one-shot structured-field transforms. `@general-paid`: quality-sensitive — offer evaluation narrative Blocks A-F, cover letters, "Why X?" answers, STAR interview stories, LinkedIn outreach. `@glm-minimal`: narrow ≤5K-input one-shot extract/classify jobs that do not need context.
-  why: GLM 5.1 doesn't discount cache reads so procedural work there costs ~10×; free-tier models handle procedural work fine empirically (`opencode/big-pickle` processed 1000+ messages at $0)
+  why: OpenCode routes all JobForge tiers through DeepSeek V4 Flash by default now; recent traces showed free OpenRouter fallbacks freezing or hitting provider balance errors during applications
 
 - [D3] Read the active mode file before dispatch. Mode files own score gates, provider fallback, portal runbooks, and output shape.
   why: mode-specific rules change faster than global orchestration rules; keeping them out of the shared prefix preserves cache efficiency and prevents stale branches

package/README.md

@@ -74,8 +74,8 @@ JobForge turns opencode into a full job search command center. Instead of manual
 | **Portal Scanner** | 45+ companies pre-configured with fuzzy dedup for reposts |
 | **Batch Processing** | Parallel evaluation with `opencode run` workers, with honest verification flagging |
 | **Pipeline Integrity** | Automated merge, dedup, status normalization, health checks |
-| **Cost-Aware Agent Routing** | Three subagents (`@general-free`, `@general-paid`, `@glm-minimal`) with per-task model tiers. On OpenCode, JobForge mixes native free models with free OpenRouter routes so the harness stays no-cost without forcing every task through the same provider. See [Subagent Routing in AGENTS.md](AGENTS.md) for the task-to-agent mapping. |
-| **Automatic Model Fallback** | When a model rate-limits or 5xx's, [`@razroo/opencode-model-fallback`](https://www.npmjs.com/package/@razroo/opencode-model-fallback) rotates the agent through a configured `fallback_models` chain and replays the request. JobForge's OpenCode defaults stay on free models for both primaries and fallbacks. |
+| **Cost-Aware Agent Routing** | Three subagents (`@general-free`, `@general-paid`, `@glm-minimal`) with per-task tool surfaces. On OpenCode, JobForge pins all tiers to `opencode-go/deepseek-v4-flash` so application runs avoid overloaded free-model pools. See [Subagent Routing in AGENTS.md](AGENTS.md) for the task-to-agent mapping. |
+| **Trace + Telemetry** | `job-forge trace:*` exposes local OpenCode transcripts, and `job-forge telemetry:*` summarizes runs, child outcomes, provider errors, and pending tracker TSVs. |
 | **Token Cost Visibility** | `job-forge tokens --days 1` for per-session breakdown; `job-forge session-report --since-minutes 60 --log` to flag sessions over budget and append history to `data/token-usage.tsv`. Auto-logged after every batch run. |
 
 ## Usage

package/bin/create-job-forge.mjs

@@ -117,22 +117,14 @@ const consumerPkg = {
     'telemetry:status': 'job-forge telemetry:status',
     'telemetry:show': 'job-forge telemetry:show',
     'telemetry:watch': 'job-forge telemetry:watch',
-    // One command to pull the latest harness, companion plugin, and any
-    // locally-pinned MCP packages. npm update is a no-op on packages not
-    // in package.json, so listing @razroo/gmail-mcp + @geometra/mcp is
-    // safe for consumers that invoke them via `npx -y` without pinning.
-    'update-harness': 'npm update job-forge @razroo/opencode-model-fallback @razroo/gmail-mcp @geometra/mcp && job-forge sync && node -e "console.log(\'✅ harness at\', require(\'./package-lock.json\').packages[\'node_modules/job-forge\'].resolved)"',
+    // One command to pull the latest harness and any locally-pinned MCP
+    // packages. npm update is a no-op on packages not in package.json, so
+    // listing @razroo/gmail-mcp + @geometra/mcp is safe for consumers that
+    // invoke them via `npx -y` without pinning.
+    'update-harness': 'npm update job-forge @razroo/gmail-mcp @geometra/mcp && job-forge sync && node -e "console.log(\'✅ harness at\', require(\'./package-lock.json\').packages[\'node_modules/job-forge\'].resolved)"',
   },
   dependencies: {
     'job-forge': '^2.0.0',
-    // Model-fallback plugin: rotates agents through their fallback_models
-    // chain on rate-limit / 5xx errors so a rate-limited free-tier model
-    // doesn't wedge the whole flow. The chains live upstream in each
-    // agent's MD frontmatter (`.opencode/agents/*.md` in the harness);
-    // consumers can override individual chains by adding their own
-    // agent.<name>.fallback_models block to opencode.json. Requires
-    // 0.3.1+ for the frontmatter-merge path.
-    '@razroo/opencode-model-fallback': '^0.3.1',
   },
   engines: { node: '>=18' },
 };
@@ -142,16 +134,11 @@ write('package.json', JSON.stringify(consumerPkg, null, 2) + '\n');
 
 const opencodeCfg = {
   $schema: 'https://opencode.ai/config.json',
-  // Keep the top-level orchestrator on a free model too. Subagents pin
-  // their own models in .opencode/agents/*.md; this covers the main chat
-  // session and any commands that don't hop to a subagent immediately.
-  model: 'openrouter/qwen/qwen3-coder:free',
-  small_model: 'openrouter/google/gemma-4-26b-a4b-it:free',
-  // Model-fallback plugin: on rate-limit / 5xx / known provider errors,
-  // rotates the agent's model to the next entry in its fallback_models
-  // chain (see `agent` below) and replays the request. Without this, a
-  // rate-limited free-tier model wedges the whole subagent flow.
-  plugin: ['@razroo/opencode-model-fallback'],
+  // Keep the top-level orchestrator on JobForge's low-cost paid OpenCode
+  // route. Subagents pin the same route in .opencode/agents/*.md so job
+  // applications do not fall through overloaded free OpenRouter pools.
+  model: 'opencode-go/deepseek-v4-flash',
+  small_model: 'opencode-go/deepseek-v4-flash',
   // Files listed here load into every session's cached prefix, so they're
   // cached once (on Anthropic) instead of Read-as-tool-call on every session.
   //   AGENTS.harness.md → symlink to node_modules/job-forge/AGENTS.md (harness rules)
@@ -182,31 +169,6 @@ const opencodeCfg = {
       environment: { DISABLE_HTTP: 'true' },
     },
   },
-  // Register the exact OpenRouter free models the harness uses so they're
-  // selectable even if they are not in OpenCode's built-in preloaded set.
-  // This list is a superset: role primaries, per-agent fallback chains,
-  // and the orchestrator fallback chain.
-  provider: {
-    openrouter: {
-      models: {
-        // Orchestrator + agentic coding (role default)
-        'qwen/qwen3-coder:free': {},
-        // Role primaries
-        'z-ai/glm-4.5-air:free': {}, // fast
-        'qwen/qwen3-next-80b-a3b-instruct:free': {}, // quality
-        'openai/gpt-oss-20b:free': {}, // minimal
-        // Common fallbacks
-        'openai/gpt-oss-120b:free': {},
-        'minimax/minimax-m2.5:free': {},
-        'nvidia/nemotron-3-super-120b-a12b:free': {},
-        'nvidia/nemotron-3-nano-30b-a3b:free': {},
-        'nvidia/nemotron-nano-9b-v2:free': {},
-        'google/gemma-4-26b-a4b-it:free': {},
-        'google/gemma-4-31b-it:free': {},
-        'meta-llama/llama-3.3-70b-instruct:free': {},
-      },
-    },
-  },
   // Restrict the primary orchestrator to dispatching only the three harness
   // subagents. Prevents accidental self-calls or unregistered agents.
   // Override locally in opencode.json if you add project-specific agents.

package/bin/sync.mjs

@@ -75,7 +75,6 @@ const links = [
 
   // OpenCode: skill router + subagent definitions. Users can override any
   // single subagent by replacing its symlink with a local file.
-  { src: '.opencode/opencode-model-fallback.json', dst: '.opencode/opencode-model-fallback.json' },
   { src: '.opencode/skills/job-forge.md',  dst: '.opencode/skills/job-forge.md' },
   { src: '.opencode/agents',               dst: '.opencode/agents' },
 

package/docs/ARCHITECTURE.md

@@ -43,11 +43,11 @@ The consumer's `opencode.json` loads a small set of stable files as always-prese
 
 The skill router (`.opencode/skills/job-forge.md`) loads mode and data files on demand, keeping per-session input tokens low (~20-40K for most modes instead of ~130-170K when everything was force-loaded).
 
-**Cost-tiered subagents** live in `.opencode/agents/` (`general-free`, `general-paid`, `glm-minimal`). On OpenCode, JobForge now uses a mix of native free models and free OpenRouter routes, with different quality/latency tiers per task shape. See [MODEL-ROUTING.md](MODEL-ROUTING.md) for the routing architecture, why it exists, and how to customize.
+**Cost-tiered subagents** live in `.opencode/agents/` (`general-free`, `general-paid`, `glm-minimal`). On OpenCode, JobForge pins all three tiers to `opencode-go/deepseek-v4-flash` by default, while the tiers still differ by tool surface, reasoning budget, and task prompt. See [MODEL-ROUTING.md](MODEL-ROUTING.md) for the routing architecture, why it exists, and how to customize.
 
 **Multi-harness support.** Because `iso/` is the single source of truth, publishing ships config for OpenCode, Cursor, Claude Code, and Codex in one tarball. Consumers run any of `opencode`, `cursor`, `claude`, or `codex` in the project and each picks up the shared MCP config + instructions via the symlinks above.
 
-**Upgrading** the harness in a consumer project is `npm run update-harness` — pulls the latest `job-forge` from npm, refreshes the fallback plugin + pinned MCPs, re-runs symlink sync, and prints the resolved version.
+**Upgrading** the harness in a consumer project is `npm run update-harness` — pulls the latest `job-forge` from npm, refreshes pinned MCPs, re-runs symlink sync, and prints the resolved version.
 
 ## System Overview
 

package/docs/MODEL-ROUTING.md

@@ -4,13 +4,13 @@ JobForge routes each piece of work to the cheapest model that can do it well, in
 
 ## Why routing matters (the cost math)
 
-A two-day trace early in development showed `$48` in spend, with **84% coming from GLM 5.1** despite the majority of the work being procedural (form fills, tracker updates, OTP retrieval). The root cause:
+A two-day trace early in development showed `$48` in spend, with **84% coming from GLM 5.1** despite the majority of the work being procedural (form fills, tracker updates, OTP retrieval). Later traces showed the opposite failure mode: free OpenRouter routes hit shared-pool contention and Venice balance errors during application subagents.
 
 - **GLM 5.1's provider doesn't discount cache reads.** On Anthropic, a 10K-token cached prefix costs ~$0.03. On GLM 5.1 it bills near-full input rate (~$0.35). Every session that re-loads the prefix pays full price.
-- **Procedural work is the high-volume work.** 1000+ messages per day go to form filling, TSV merges, scan dedup. Running that on a paid model is unnecessary when current free models can handle the task.
-- **OpenCode no longer needs one provider for every role.** JobForge now pins the procedural `@general-free` worker to `opencode/big-pickle`, while the quality-sensitive writer tier stays on a free OpenRouter route.
+- **Procedural work is the high-volume work.** 1000+ messages per day go to form filling, TSV merges, scan dedup. It needs a cheap reliable route, not the absolute strongest model.
+- **Free pools are not reliable enough for application runs.** A 2026-04-25 trace showed `@general-paid` hitting `openrouter/qwen/...:free` Venice 402 errors, while `@general-free` fell through `opencode/big-pickle` into `z-ai/glm-4.5-air:free` and `gpt-oss-20b:free`.
 
-Conclusion: route procedural work to free tier, reserve paid models for tasks that actually need the quality.
+Conclusion: keep the subagent split for tool permissions and prompting, but route OpenCode to `opencode-go/deepseek-v4-flash` across all JobForge tiers. It is the best current "affordable and reliable" default for applications.
 
 ## The three subagents
 
@@ -18,9 +18,9 @@ Defined in `.opencode/agents/*.md` (shipped in the harness, symlinked into consu
 
 | Agent | Model | Reasoning | Use for |
 |-------|-------|-----------|---------|
-| `@general-free` | `opencode/big-pickle` | `minimal` | Geometra form fills, tracker TSV merges, scan dedup, OTP retrieval via Gmail, scripted pipeline steps |
-| `@general-paid` | `openrouter/qwen/qwen3-next-80b-a3b-instruct:free` | `medium` | Offer evaluation narratives (Blocks A-F), cover letters, "Why X?" answers, STAR+R interview stories, LinkedIn outreach prose |
-| `@glm-minimal` | `openrouter/openai/gpt-oss-20b:free` | `none` | Narrow one-shot transforms: "extract these 8 fields from this JD text → JSON", "classify this archetype" |
+| `@general-free` | `opencode-go/deepseek-v4-flash` | `minimal` | Geometra form fills, tracker TSV merges, scan dedup, OTP retrieval via Gmail, scripted pipeline steps |
+| `@general-paid` | `opencode-go/deepseek-v4-flash` | `medium` | Offer evaluation narratives (Blocks A-F), cover letters, "Why X?" answers, STAR+R interview stories, LinkedIn outreach prose |
+| `@glm-minimal` | `opencode-go/deepseek-v4-flash` | `none` | Narrow one-shot transforms: "extract these 8 fields from this JD text → JSON", "classify this archetype" |
 
 The full task-to-agent mapping lives in [AGENTS.md → Subagent Routing](../AGENTS.md#subagent-routing--which-agent-for-which-task). The orchestrator (your primary session) is expected to delegate before taking any multi-step action — see the **Pre-flight delegation** rule in AGENTS.md.
 
@@ -64,25 +64,35 @@ Disables ~30 MCP tool schemas globally; each agent re-enables only what it needs
 
 All three layers are designed to be edited — this is your search, your cost budget, your model preferences.
 
-### Swap the paid model
+### Swap the quality model
 
-The default `@general-paid` is `openrouter/qwen/qwen3-next-80b-a3b-instruct:free`. To use Claude instead, edit `.opencode/agents/general-paid.md`:
+The default `@general-paid` OpenCode model is `opencode-go/deepseek-v4-flash`. To use a stronger OpenCode route for quality writing, edit `models.yaml`:
 
 ```yaml
----
-model: opencode/claude-sonnet-4-6
-reasoningEffort: medium
----
+roles:
+  quality:
+    targets:
+      opencode:
+        provider: opencode
+        model: opencode-go/deepseek-v4-pro
 ```
 
-The `.opencode/agents/general-paid.md` file is a symlink into `node_modules/job-forge/` by default. To customize locally without modifying the harness: delete the symlink and create a real file with the same name — `job-forge sync` will skip it on future updates. Or override in `opencode.json` under `agent.general-paid.model`.
+Run `npm run build:config` if you are editing a harness checkout. In a consumer project, keep local overrides in `opencode.json` or replace a symlinked `.opencode/agents/<name>.md` with a real file.
 
-### Swap the free tier
+### Swap the procedural route
 
-The primary `@general-free` model is set in `models.yaml` via the `fast`
-role's `targets.opencode` override. Change that if you want a different
-default. The `fallback_models` list still lives in
-`.opencode/agents/general-free.md` / `iso/agents/general-free.md`.
+The `@general-free` model is set in `models.yaml` via the `fast` role's
+`targets.opencode` override. Change that if you want a different procedural
+default:
+
+```yaml
+roles:
+  fast:
+    targets:
+      opencode:
+        provider: opencode
+        model: opencode-go/deepseek-v4-flash
+```
 
 ### Add a custom agent
 
@@ -125,8 +135,8 @@ npx job-forge tokens --session <session-id>
 ```
 
 Healthy pattern after this architecture lands:
-- **Free OpenRouter defaults** now cover procedural, quality-sensitive, and extractor work on OpenCode
-- If you opt back into a paid model, do it deliberately and only for the role that needs it
+- **DeepSeek V4 Flash defaults** cover procedural, quality-sensitive, and extractor work on OpenCode
+- If you opt into a stronger model, do it deliberately and only for the role that needs it
 - Session titles prefixed `@general-free`, `@general-paid`, `@glm-minimal` appear in the list — confirms delegation actually happened
 - `cache_read` >> `cache_creation` on parallel subagent runs within a 5-min window
 
@@ -134,67 +144,24 @@ Failure pattern to watch for:
 - **All messages showing up under your primary model** (no `@general-*` titles) → orchestrator isn't delegating. Check the Pre-flight delegation rule in AGENTS.md is being followed; tighten wording if not.
 - **High cache-creation with near-zero cache-read across parallel workers** → workers aren't firing within the 5-min cache TTL, or the shared prefix isn't byte-identical (see [batch/README.md](../batch/README.md) for the prompt-cache-friendly batch pattern).
 
-## Automatic fallback on rate limits / 5xx
-
-A rate-limited or overloaded free-tier model would otherwise wedge the whole subagent flow — the delegated task errors and the orchestrator sits stuck. The harness ships with [`@razroo/opencode-model-fallback`](https://www.npmjs.com/package/@razroo/opencode-model-fallback) (added as a dependency in the scaffolder) to rotate agents through a configured `fallback_models` chain automatically.
-
-Default chains ship upstream in each agent's YAML frontmatter (`node_modules/job-forge/.opencode/agents/*.md`, symlinked into your project's `.opencode/agents/`):
-
-| Agent | Primary | Fallback chain (in order) |
-|-------|---------|---------------------------|
-| `@general-free` | `opencode/big-pickle` | `openrouter/z-ai/glm-4.5-air:free` → `openrouter/openai/gpt-oss-20b:free` → `openrouter/nvidia/nemotron-3-nano-30b-a3b:free` → `openrouter/qwen/qwen3-coder:free` |
-| `@general-paid` | `openrouter/qwen/qwen3-next-80b-a3b-instruct:free` | `openrouter/openai/gpt-oss-120b:free` → `openrouter/nvidia/nemotron-3-super-120b-a12b:free` → `openrouter/z-ai/glm-4.5-air:free` → `openrouter/qwen/qwen3-coder:free` → `openrouter/google/gemma-4-31b-it:free` → `openrouter/meta-llama/llama-3.3-70b-instruct:free` |
-| `@glm-minimal` | `openrouter/openai/gpt-oss-20b:free` | `openrouter/google/gemma-4-26b-a4b-it:free` → `openrouter/nvidia/nemotron-nano-9b-v2:free` → `openrouter/google/gemma-4-31b-it:free` → `openrouter/z-ai/glm-4.5-air:free` |
-
-These chains are deliberately free-only so the default OpenCode path never needs to pay. **Note:** OpenCode model IDs must use the provider prefix it expects (`openrouter/...`, `opencode/...`, etc.). The raw OpenRouter model slug by itself is not enough.
-
-Consumers **do not need to configure anything** to get these defaults: the subagent chains arrive via the symlinked agent MD files, and the harness ships `.opencode/opencode-model-fallback.json` for the main orchestrator / any agent without its own list. That JSON is **generated by `@razroo/iso-harness`** from the top-level `opencodeModelFallback` object in the harness's `iso/config.json` (same `iso build` pass as `opencode.json`), so it stays in sync with published releases — not a hand-edited file in git. `@razroo/opencode-model-fallback` (≥0.3.1) reads per-agent chains from the frontmatter-derived `fallback_models` field and falls through to the global file when no per-agent list exists. The consumer's `opencode.json` only needs `"plugin": ["@razroo/opencode-model-fallback"]` — which the scaffolder sets automatically.
-
-**When fallback fires:** the plugin pattern-matches rate-limit / 5xx / quota / "overloaded" / insufficient-balance style errors (including Venice / Diem copy — see `retryable_error_patterns` in the harness source). Failed models enter a cooldown before they're retried. Every rotation logs to `~/.config/opencode/opencode-model-fallback.log` with the trigger error, original model, and target model — grep for `"Auto-retrying with fallback model"` to confirm it fired.
-
-**Orchestrator policy (all harnesses):** if an upgraded `@general-paid` apply still fails after model fallback, follow **[D3f]** in the shared instructions (`AGENTS.md` / `CLAUDE.md` / Cursor rules / Codex `AGENTS.md`) — re-dispatch the same URL once on `@general-free` before marking FAILED.
-
-### Overriding an upstream chain
-
-Add an `agent.<name>.fallback_models` block to your project's `opencode.json`. Top-level entries win over upstream frontmatter:
-
-```json
-{
-  "agent": {
-    "general-free": {
-      "fallback_models": ["my/preferred-free", "my/preferred-paid"]
-    }
-  }
-}
-```
-
-### Global fallback chain (agents without their own)
-
-Source of truth: **`iso/config.json` → `opencodeModelFallback`** (iso-harness writes `.opencode/opencode-model-fallback.json` on `npm run build:config` / `prepack`). Applies to the orchestrator and any agent whose `fallback_models` is empty. Example shape:
-
-```json
-{
-  "cooldown_seconds": 60,
-  "timeout_seconds": 30,
-  "notify_on_fallback": true,
-  "fallback_models": [
-    "openrouter/openai/gpt-oss-120b:free",
-    "openrouter/z-ai/glm-4.5-air:free"
-  ],
-  "retryable_error_patterns": ["(?i)venice.*insufficient"]
-}
-```
-
-`retryable_error_patterns` extends the plugin's built-in matchers — use it for provider-specific strings (Venice Diem balance, etc.) that do not match the stock regexes.
+## Fallback policy
 
-### Disabling fallback
+JobForge no longer ships automatic free-model fallback for OpenCode. The
+previous free fallback chain solved some transient model outages, but real
+application traces showed it could also freeze runs upstream or route into
+provider balance failures. The default is now simpler: each OpenCode role uses
+`opencode-go/deepseek-v4-flash`, and errors surface directly in telemetry.
 
-Remove `"@razroo/opencode-model-fallback"` from `opencode.json:plugin` — agents keep their `model:` primary and errors propagate normally.
+Use `npm run telemetry:status` or `npm run telemetry:show -- <session>` to
+inspect provider errors, child outcomes, and pending TSVs after a run. If you
+want a local fallback chain anyway, add it explicitly in your own
+`opencode.json` or agent frontmatter so the cost/reliability tradeoff is
+visible in your project rather than hidden in the harness defaults.
 
 ## Known limitations
 
 - **opencode's 5-minute cache TTL is hardcoded.** The 1-hour cache (Anthropic beta, `extended-cache-ttl-2025-04-11`) is not plumbed through opencode as of 2026-04-15. Long batch runs (>5 min between workers) will miss cache every cycle. Upstream fix would be 2 lines in `packages/opencode/src/provider/`.
-- **`instructions` is top-level, not per-agent.** Files listed in `opencode.json:instructions` load for every agent including free-tier. This is fine for `cv.md` and `_shared.md` (they're small and useful everywhere), but means you can't hide heavy context from free agents via instructions — use per-agent `prompt:` files for that.
+- **`instructions` is top-level, not per-agent.** Files listed in `opencode.json:instructions` load for every agent. This is fine for `cv.md` and `_shared.md` (they're small and useful everywhere), but means you can't hide heavy context from lower-cost agents via instructions — use per-agent `prompt:` files for that.
 - **`reasoningEffort` support varies by provider.** Anthropic accepts `thinking: { type: "disabled" }`; opencode-labs models may need the `variant` pattern. See the `reasoningEffort` values in the opencode docs.
 
 ## See also

package/docs/README.md

@@ -20,7 +20,7 @@ See [SETUP.md](SETUP.md) for both paths.
 | [SETUP.md](SETUP.md) | Prerequisites, both install paths, profile and CV, portals, `npx job-forge verify`, optional Go dashboard, token usage tracking, [troubleshooting](SETUP.md#troubleshooting) |
 | [ARCHITECTURE.md](ARCHITECTURE.md) | [Package architecture](ARCHITECTURE.md#package-architecture-v200) (consumer vs harness split), modes under `modes/`, single-offer flow, batch runner, tracker and scripts, [contributor touchpoints](ARCHITECTURE.md#contributing-touchpoints) |
 | [CUSTOMIZATION.md](CUSTOMIZATION.md) | Profile, archetypes in `_shared.md`, `portals.yml`, CV template, canonical states, optional story bank and opencode hooks. Also: how to customize a symlinked mode file locally |
-| [MODEL-ROUTING.md](MODEL-ROUTING.md) | Why the harness uses three cost-tiered subagents (`@general-free`, `@general-paid`, `@glm-minimal`), how routing is enforced (permission + tool-surface trim + reasoningEffort), how to swap models or add agents, and the automatic fallback-on-rate-limit behavior via [`@razroo/opencode-model-fallback`](https://www.npmjs.com/package/@razroo/opencode-model-fallback) |
+| [MODEL-ROUTING.md](MODEL-ROUTING.md) | Why the harness uses three cost-tiered subagents (`@general-free`, `@general-paid`, `@glm-minimal`), how routing is enforced (permission + tool-surface trim + reasoningEffort), and how to swap models or add agents |
 | [examples/README.md](../examples/README.md) | Fictional CV samples (per-role markdown), optional digest, fictional [`sample-jd.md`](../examples/sample-jd.md) for `local:jds/…` shape, and sample report — starting point for new `cv.md`, JD-on-disk layout, or contributor archetypes (cloud infrastructure, agent platform) |
 | [batch/README.md](../batch/README.md) | Batch TSV format, `batch-runner.sh`, `tracker-additions/` merge flow with `npx job-forge merge` |
 | [jds/README.md](../jds/README.md) | Markdown JDs on disk; `local:jds/{file}.md` lines in `data/pipeline.md` |

package/iso/agents/general-free.md

@@ -1,24 +1,16 @@
 ---
-description: Procedural worker on free-tier model. Use for form filling via Geometra, tracker updates, TSV merges, scan dedup, OTP retrieval, and other mechanical/scripted tasks where quality-sensitive text generation is NOT required.
+description: Procedural worker on the low-cost DeepSeek V4 Flash OpenCode route. Use for form filling via Geometra, tracker updates, TSV merges, scan dedup, OTP retrieval, and other mechanical/scripted tasks where quality-sensitive text generation is NOT required.
 role: fast
 targets:
-  # No inline model: iso-route's "standard" preset maps role "fast" to
-  # each harness's cheapest reliable model. Claude Code reads
+  # No inline model: JobForge's models.yaml maps role "fast" to each
+  # harness's cheapest reliable model. On OpenCode that is pinned to
+  # opencode-go/deepseek-v4-flash. Claude Code reads
   # .claude/iso-route.resolved.json; OpenCode reads opencode.json's
   # agent.fast.model (iso-harness 0.6.0+).
   opencode:
     mode: subagent
     temperature: 0.1
     reasoningEffort: minimal
-    # Primary comes from models.yaml: opencode/big-pickle on OpenCode.
-    # Fallback chain stays free-only and intentionally excludes
-    # openrouter/minimax/minimax-m2.5:free because recent traces showed
-    # repeated read({ path|file_path }) schema drift on that route.
-    fallback_models:
-      - openrouter/z-ai/glm-4.5-air:free
-      - openrouter/openai/gpt-oss-20b:free
-      - openrouter/nvidia/nemotron-3-nano-30b-a3b:free
-      - openrouter/qwen/qwen3-coder:free
     tools:
       geometra_connect: true
       geometra_page_model: true
@@ -34,7 +26,7 @@ targets:
       task: false
 ---
 
-You are the @general-free subagent. You run on a free-tier model, which means the orchestrator has delegated this task to you **specifically because the work is procedural**: deterministic steps, scripted outputs, no nuanced writing required.
+You are the @general-free subagent. You run on JobForge's low-cost procedural model, which means the orchestrator has delegated this task to you **specifically because the work is procedural**: deterministic steps, scripted outputs, no nuanced writing required.
 
 ## Run This Pre-Flight First Every Time
 

package/iso/agents/general-paid.md

@@ -1,28 +1,16 @@
 ---
-description: Quality-sensitive worker on the strongest free-tier OpenCode model by default. Use for offer evaluation narratives (Blocks A-F), cover letter generation, "Why X?" form answers, interview STAR stories, and other tasks where writing quality and judgment matter.
+description: Quality-sensitive worker on the low-cost DeepSeek V4 Flash OpenCode route by default. Use for offer evaluation narratives (Blocks A-F), cover letter generation, "Why X?" form answers, interview STAR stories, and other tasks where writing quality and judgment matter.
 role: quality
 targets:
-  # No inline model: JobForge's models.yaml maps role "quality" to a
-  # free OpenRouter model on OpenCode, while Claude/Codex keep their
-  # quality-tier defaults from the standard preset. Claude Code reads
+  # No inline model: JobForge's models.yaml maps role "quality" to
+  # opencode-go/deepseek-v4-flash on OpenCode, while Claude/Codex keep
+  # their quality-tier defaults from the standard preset. Claude Code reads
   # .claude/iso-route.resolved.json; OpenCode reads opencode.json's
   # agent.quality.model (iso-harness 0.6.0+).
   opencode:
     mode: subagent
     temperature: 0.3
     reasoningEffort: medium
-    # Primary (qwen/qwen3-next-80b-a3b-instruct:free) resolves from the
-    # openrouter-free preset. First fallbacks intentionally avoid another
-    # immediate hop through the same Venice/Qwen pool when OpenRouter
-    # returns "[Venice] insufficient …" — gpt-oss-120b + nemotron are
-    # usually different backends. Remaining picks stay free-only.
-    fallback_models:
-      - openrouter/openai/gpt-oss-120b:free
-      - openrouter/nvidia/nemotron-3-super-120b-a12b:free
-      - openrouter/z-ai/glm-4.5-air:free
-      - openrouter/qwen/qwen3-coder:free
-      - openrouter/google/gemma-4-31b-it:free
-      - openrouter/meta-llama/llama-3.3-70b-instruct:free
     tools:
       geometra_connect: true
       geometra_page_model: true
@@ -40,10 +28,10 @@ targets:
 
 You are the @general-paid subagent. The orchestrator delegated this task to you because it requires quality writing or judgment — the kind of work `@general-free` isn't well-suited for.
 
-On OpenCode, this agent now defaults to a free OpenRouter model. On other
-harnesses, the same role may still resolve to a premium model. Your job is
-still the same: produce the best final writing you can from the context you
-were given.
+On OpenCode, this agent defaults to DeepSeek V4 Flash so application work
+does not fall back into overloaded free OpenRouter pools. On other harnesses,
+the same role may still resolve to a premium model. Your job is still the
+same: produce the best final writing you can from the context you were given.
 
 ## Do These Tasks
 

package/iso/agents/glm-minimal.md

@@ -1,23 +1,16 @@
 ---
-description: Narrow-scope extractor on free-tier model. Use for single-purpose tasks where the orchestrator passes the exact input and expects a small, structured output — e.g., "extract these 8 fields from this JD text" or "parse this form schema into a label→type map". NOT for multi-step workflows.
+description: Narrow-scope extractor on the low-cost DeepSeek V4 Flash OpenCode route. Use for single-purpose tasks where the orchestrator passes the exact input and expects a small, structured output — e.g., "extract these 8 fields from this JD text" or "parse this form schema into a label→type map". NOT for multi-step workflows.
 role: minimal
 targets:
-  # No inline model: iso-route's "standard" preset maps role "minimal" to
-  # each harness's smallest credible model. Claude Code reads
+  # No inline model: JobForge's models.yaml maps role "minimal" to each
+  # harness's smallest credible model. On OpenCode that is pinned to
+  # opencode-go/deepseek-v4-flash. Claude Code reads
   # .claude/iso-route.resolved.json; OpenCode reads opencode.json's
   # agent.minimal.model (iso-harness 0.6.0+).
   opencode:
     mode: subagent
     temperature: 0
     reasoningEffort: none
-    # Primary (openai/gpt-oss-20b:free) resolves from openrouter-free
-    # preset. Fallback chain sticks to small dense models with reliable
-    # structured-output behavior — no creative generation upstream.
-    fallback_models:
-      - openrouter/google/gemma-4-26b-a4b-it:free
-      - openrouter/nvidia/nemotron-nano-9b-v2:free
-      - openrouter/google/gemma-4-31b-it:free
-      - openrouter/z-ai/glm-4.5-air:free
     tools:
       geometra_*: false
       gmail_*: false

package/iso/config.json

@@ -1,49 +1,8 @@
 {
-  "opencodeModelFallback": {
-    "cooldown_seconds": 60,
-    "timeout_seconds": 30,
-    "notify_on_fallback": true,
-    "fallback_models": [
-      "openrouter/openai/gpt-oss-120b:free",
-      "openrouter/z-ai/glm-4.5-air:free",
-      "openrouter/nvidia/nemotron-3-super-120b-a12b:free",
-      "openrouter/qwen/qwen3-next-80b-a3b-instruct:free"
-    ],
-    "retryable_error_patterns": [
-      "\\bvenice\\b",
-      "insufficient\\s+usd",
-      "insufficient\\s+.*\\s+diem",
-      "diem\\s+balance",
-      "add\\s+credits",
-      "chutes",
-      "insufficient\\s+(?:credits?|funds?|balance)",
-      "credit.*balance.*too.*low",
-      "(?:temporarily\\s+)?unavailable|overloaded|try\\s+again"
-    ]
-  },
   "targets": {
     "opencode": {
-      "plugin": ["@razroo/opencode-model-fallback"],
       "instructions": ["templates/states.yml"],
-      "small_model": "openrouter/google/gemma-4-26b-a4b-it:free",
-      "provider": {
-        "openrouter": {
-          "models": {
-            "qwen/qwen3-coder:free": {},
-            "z-ai/glm-4.5-air:free": {},
-            "qwen/qwen3-next-80b-a3b-instruct:free": {},
-            "openai/gpt-oss-20b:free": {},
-            "openai/gpt-oss-120b:free": {},
-            "minimax/minimax-m2.5:free": {},
-            "nvidia/nemotron-3-super-120b-a12b:free": {},
-            "nvidia/nemotron-3-nano-30b-a3b:free": {},
-            "nvidia/nemotron-nano-9b-v2:free": {},
-            "google/gemma-4-26b-a4b-it:free": {},
-            "google/gemma-4-31b-it:free": {},
-            "meta-llama/llama-3.3-70b-instruct:free": {}
-          }
-        }
-      }
+      "small_model": "opencode-go/deepseek-v4-flash"
     }
   }
 }

package/iso/instructions.md

@@ -5,7 +5,7 @@ AI-powered job search pipeline: scans portals, evaluates offers, generates CVs v
 ## Hard limits
 
 - [H1] Max 2 parallel `task` dispatches per message. For N jobs, run `ceil(N/2)` sequential rounds of 2. A round is not complete until both subagents return a final outcome (`APPLIED`, `APPLY FAILED`, `SKIP`, `Discarded`, or a written TSV path). A `task` tool result that only gives a session id / title is a launch acknowledgement, not completion. Applies in all modes, for all user phrasings ("urgent", "apply to 10 jobs now").
-  why: higher parallelism blows through free-tier rate limits; each subagent requires post-cleanup and racing more than 2 reliably loses at least one result. On 2026-04-25 the orchestrator launched round 2 while round 1 had only returned task ids, leaving four application subagents in flight and losing two provider-fallback recoveries
+  why: each subagent requires post-cleanup and racing more than 2 reliably loses at least one result. On 2026-04-25 the orchestrator launched round 2 while round 1 had only returned task ids, leaving four application subagents in flight and losing two provider recoveries
 
 - [H2] Max 1 application per company+role. Before every `apply` dispatch, grep all four sources for the URL and for `company+role`: `data/pipeline.md`, all `data/applications/*.md` day files, `batch/tracker-additions/*.tsv`, `batch/tracker-additions/merged/*.tsv`. If any source shows APPLIED / Applied, skip the dispatch.
   why: 2026-04 same-day batch collision — when two batches target the same role, `npx job-forge merge` updates the existing day-file row rather than appending, so grepping day files alone misses earlier-batch applies; merged/*.tsv is the only place the breadcrumb remains
@@ -37,7 +37,7 @@ AI-powered job search pipeline: scans portals, evaluates offers, generates CVs v
   why: iso-trace showed 0.25% Agent calls across 5174 turns under a prior over-broad "delegate before 2nd tool call" rule — the rule was ignored in practice; narrowing matches the original cache-bust incident
 
 - [D2] Route subagent work by cost tier. `@general-free`: procedural — form-fill, TSV merge, verify, OTP retrieval, portal scan metadata extraction, one-shot structured-field transforms. `@general-paid`: quality-sensitive — offer evaluation narrative Blocks A-F, cover letters, "Why X?" answers, STAR interview stories, LinkedIn outreach. `@glm-minimal`: narrow ≤5K-input one-shot extract/classify jobs that do not need context.
-  why: GLM 5.1 doesn't discount cache reads so procedural work there costs ~10×; free-tier models handle procedural work fine empirically (`opencode/big-pickle` processed 1000+ messages at $0)
+  why: OpenCode routes all JobForge tiers through DeepSeek V4 Flash by default now; recent traces showed free OpenRouter fallbacks freezing or hitting provider balance errors during applications
 
 - [D3] Read the active mode file before dispatch. Mode files own score gates, provider fallback, portal runbooks, and output shape.
   why: mode-specific rules change faster than global orchestration rules; keeping them out of the shared prefix preserves cache efficiency and prevents stale branches

package/models.yaml

@@ -1,17 +1,15 @@
 # JobForge model policy.
 #
-# Extends @razroo/iso-route's bundled "openrouter-free" preset — a
-# curated zero-cost routing config that keeps Claude Code / Codex on
-# their native providers but routes OpenCode through explicit free
-# OpenRouter model IDs. Preset content lives in
-# node_modules/@razroo/iso-route/presets/openrouter-free.yaml; run
-# `npx iso-route plan models.yaml` to see the resolved policy.
+# Extends @razroo/iso-route's bundled "standard" preset, then pins every
+# OpenCode route to DeepSeek V4 Flash. Recent traces showed free OpenRouter
+# routes freezing or falling through Venice balance errors, so JobForge's
+# OpenCode default is now "best affordable paid" rather than "free".
 #
 # JobForge's subagents bind to preset roles via the `role:` field in
 # iso/agents/<slug>.md:
-#   @general-free  → role: fast     (Haiku / OpenCode big-pickle / gpt-5.4-mini)
-#   @general-paid  → role: quality  (Opus 4.7 / OpenRouter Qwen3 Next 80B free / gpt-5.4)
-#   @glm-minimal   → role: minimal  (Haiku / OpenRouter GPT-OSS-20B free / gpt-5.4-nano)
+#   @general-free  → role: fast     (Haiku / DeepSeek V4 Flash / gpt-5.4-mini)
+#   @general-paid  → role: quality  (Opus 4.7 / DeepSeek V4 Flash / gpt-5.4)
+#   @glm-minimal   → role: minimal  (Haiku / DeepSeek V4 Flash / gpt-5.4-nano)
 #
 # Override anything by adding fields here. For example, to pin Opus on
 # Claude Code for the @general-paid (quality) role:
@@ -29,11 +27,27 @@
 #           provider: openai
 #           model: gpt-5.4
 
-extends: openrouter-free
+extends: standard
+
+default:
+  targets:
+    opencode:
+      provider: opencode
+      model: opencode-go/deepseek-v4-flash
 
 roles:
   fast:
     targets:
       opencode:
         provider: opencode
-        model: opencode/big-pickle
+        model: opencode-go/deepseek-v4-flash
+  quality:
+    targets:
+      opencode:
+        provider: opencode
+        model: opencode-go/deepseek-v4-flash
+  minimal:
+    targets:
+      opencode:
+        provider: opencode
+        model: opencode-go/deepseek-v4-flash

package/modes/apply.md

@@ -17,7 +17,7 @@ Live application assistant. Reads the active application form in Chrome (via Geo
   why: prior aborted subagents leave Chromium sessions stuck in the pool; next `geometra_connect` fails with "Not connected" (see root `[H3]`)
 
 - [H5] Max 2 parallel `task` dispatches per round. For N jobs, run `ceil(N/2)` sequential rounds of 2. Never emit 3+ dispatches in a single message. Do not start the next round until both current-round subagents return final outcomes (`APPLIED`, `APPLY FAILED`, `SKIP`, `Discarded`, or a written TSV path); task/session ids are only launch receipts.
-  why: free-tier rate limits + subagent post-cleanup cost; racing more than 2 reliably loses at least one result (see root `[H1]`). A 2026-04-25 OpenCode trace launched round 2 while round 1 was still running, then lost two fallback recoveries
+  why: subagent post-cleanup cost and portal state make racing more than 2 unreliable (see root `[H1]`). A 2026-04-25 OpenCode trace launched round 2 while round 1 was still running, then lost two provider recoveries
 
 ## Defaults
 
@@ -46,10 +46,10 @@ Live application assistant. Reads the active application form in Chrome (via Geo
   why: class-B Ashby / Cloudflare-fronted portals need a residential outbound IP; the fix is wired in Geometra MCP v1.59.0 but the orchestrator owns the config pipe. See "BYO Residential Proxy" in modes/reference-portals.md.
 
 - [D8] Upgrade application routing to `@general-paid` when the offer score is ≥ 4.0/5, the user flags "top-tier", "dream job", or "high-stakes", or the candidate is late-stage/post-screen.
-  why: form-fill flows are 6+ steps; free-tier sometimes aborts mid-flow on large Greenhouse/Workday schemas; paid tier has more headroom
+  why: high-stakes applications need the quality-sensitive prompt and medium reasoning budget even though OpenCode now routes both application tiers through DeepSeek V4 Flash by default
 
-- [D9] If an upgraded `@general-paid` subagent fails with provider-side errors, re-dispatch the same URL once on `@general-free` before marking FAILED. Provider-side errors include Venice, Diem, Chutes, HTTP 402/429, insufficient credits/funds/balance, overload, and temporarily unavailable.
-  why: OpenCode paid-tier routing can still use free OpenRouter model IDs; backend pool limits are not evidence that a procedural free-tier worker cannot complete the same form after preflight gates pass
+- [D9] If a subagent fails with provider-side errors, do not auto-downgrade or re-dispatch the same URL. Report the provider failure, leave any TSV untouched unless there is a confirmed outcome, and inspect telemetry before retrying.
+  why: OpenCode now pins all JobForge application tiers to DeepSeek V4 Flash; switching `@general-paid` → `@general-free` changes the prompt/tool budget but not the provider route, so automatic duplicate dispatches add risk without fixing provider availability
 
 ## Procedure
 
@@ -65,7 +65,7 @@ Live application assistant. Reads the active application form in Chrome (via Geo
 10. Generate answers from Block B + Block F + Section G + JD.
 11. Submit as ONE `run_actions` call [H1] using labels [D6] with `imeFriendly: true` [D4].
 12. On session error, run the 4-step recovery; only one retry [H2].
-13. On upgraded-provider failure, downgrade once to `@general-free` [D9].
+13. On provider failure, stop and inspect telemetry before any retry [D9].
 14. On OTP prompt, fetch the code from Gmail via `gmail_get_message`.
 15. Submit the OTP with `geometra_fill_otp` and click Submit.
 16. Write outcome as `batch/tracker-additions/*.tsv` [H3].
@@ -98,7 +98,7 @@ Or, on failure:
 APPLY FAILED AFTER RECOVERY: <url>
   Error 1: <first error>
   Error 2: <post-recovery error>
-  Recommend: re-dispatch on @general-paid
+  Recommend: inspect telemetry before retrying this URL
 ```
 
 ---
@@ -359,10 +359,10 @@ Call 4:  geometra_run_actions({
    APPLY FAILED AFTER RECOVERY: <URL>
    Error 1: <first error message>
    Error 2: <error after recovery>
-   Recommend: re-dispatch on @general-paid
+   Recommend: inspect telemetry before retrying this URL
    ```
 
-   Do NOT try a third time. Do NOT try a different approach. The orchestrator will decide whether to re-dispatch on a bigger model.
+   Do NOT try a third time. Do NOT try a different approach. The orchestrator will decide whether to retry after inspecting telemetry.
 
 ### Skip schema re-fetches mid-flow (Rule D)
 

package/opencode.json

@@ -1,33 +1,15 @@
 {
   "$schema": "https://opencode.ai/config.json",
-  "model": "openrouter/qwen/qwen3-coder:free",
+  "model": "opencode-go/deepseek-v4-flash",
   "agent": {
     "fast": {
-      "model": "opencode/big-pickle"
+      "model": "opencode-go/deepseek-v4-flash"
     },
     "quality": {
-      "model": "openrouter/qwen/qwen3-next-80b-a3b-instruct:free"
+      "model": "opencode-go/deepseek-v4-flash"
     },
     "minimal": {
-      "model": "openrouter/openai/gpt-oss-20b:free"
-    }
-  },
-  "provider": {
-    "openrouter": {
-      "models": {
-        "qwen/qwen3-coder:free": {},
-        "z-ai/glm-4.5-air:free": {},
-        "qwen/qwen3-next-80b-a3b-instruct:free": {},
-        "openai/gpt-oss-20b:free": {},
-        "openai/gpt-oss-120b:free": {},
-        "minimax/minimax-m2.5:free": {},
-        "nvidia/nemotron-3-super-120b-a12b:free": {},
-        "nvidia/nemotron-3-nano-30b-a3b:free": {},
-        "nvidia/nemotron-nano-9b-v2:free": {},
-        "google/gemma-4-26b-a4b-it:free": {},
-        "google/gemma-4-31b-it:free": {},
-        "meta-llama/llama-3.3-70b-instruct:free": {}
-      }
+      "model": "opencode-go/deepseek-v4-flash"
     }
   },
   "mcp": {
@@ -66,11 +48,8 @@
       }
     }
   },
-  "plugin": [
-    "@razroo/opencode-model-fallback"
-  ],
   "instructions": [
     "templates/states.yml"
   ],
-  "small_model": "openrouter/google/gemma-4-26b-a4b-it:free"
+  "small_model": "opencode-go/deepseek-v4-flash"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "job-forge",
-  "version": "2.14.11",
+  "version": "2.14.12",
   "description": "AI-powered job search pipeline built on opencode",
   "type": "module",
   "bin": {
@@ -97,7 +97,6 @@
     "@razroo/iso": "^0.2.5",
     "@razroo/iso-eval": "^0.4.0",
     "@razroo/iso-harness": "^0.6.1",
-    "@razroo/iso-route": "^0.5.3",
-    "@razroo/opencode-model-fallback": "^0.3.1"
+    "@razroo/iso-route": "^0.5.3"
   }
 }

package/scripts/check-iso-smoke.mjs

@@ -21,9 +21,9 @@ const checks = [
   ["H7 distrusts subagent prose", () => every(files.instructions, ["must originate from a file", "not from prior subagent prose"])],
   ["shared prompt points to on-demand references", () => every(files.instructions, ["modes/{mode}.md", "modes/reference-setup.md", "modes/reference-portals.md", "modes/reference-geometra.md"])],
   ["apply mode owns high-stakes upgrade", () => every(files.apply, ["[D8]", "@general-paid", "4.0/5", "high-stakes"])],
-  ["apply mode owns provider downgrade", () => every(files.apply, ["[D9]", "@general-free", "HTTP 402/429", "insufficient credits/funds/balance"])],
-  ["models policy extends free OpenRouter preset", () => /extends:\s*openrouter-free/.test(files.models)],
-  ["OpenCode fallback plugin is configured", () => every(files.config, ["opencodeModelFallback", "@razroo/opencode-model-fallback"])],
+  ["apply mode blocks provider auto-downgrade", () => every(files.apply, ["[D9]", "do not auto-downgrade", "inspect telemetry before retrying"])],
+  ["models policy pins OpenCode to DeepSeek V4 Flash", () => /extends:\s*standard/.test(files.models) && count(files.models, "opencode-go/deepseek-v4-flash") >= 4],
+  ["OpenCode fallback plugin is not configured", () => !every(files.config, ["opencodeModelFallback", "@razroo/opencode-model-fallback"])],
 ];
 
 const failures = checks
@@ -41,3 +41,7 @@ console.log(`JobForge iso smoke passed (${checks.length} checks).`);
 function every(source, needles) {
   return needles.every((needle) => source.includes(needle));
 }
+
+function count(source, needle) {
+  return source.split(needle).length - 1;
+}

package/.opencode/opencode-model-fallback.json

@@ -1,22 +0,0 @@
-{
-  "cooldown_seconds": 60,
-  "timeout_seconds": 30,
-  "notify_on_fallback": true,
-  "fallback_models": [
-    "openrouter/openai/gpt-oss-120b:free",
-    "openrouter/z-ai/glm-4.5-air:free",
-    "openrouter/nvidia/nemotron-3-super-120b-a12b:free",
-    "openrouter/qwen/qwen3-next-80b-a3b-instruct:free"
-  ],
-  "retryable_error_patterns": [
-    "\\bvenice\\b",
-    "insufficient\\s+usd",
-    "insufficient\\s+.*\\s+diem",
-    "diem\\s+balance",
-    "add\\s+credits",
-    "chutes",
-    "insufficient\\s+(?:credits?|funds?|balance)",
-    "credit.*balance.*too.*low",
-    "(?:temporarily\\s+)?unavailable|overloaded|try\\s+again"
-  ]
-}