job-forge 2.0.0

package/modes/batch.md

@@ -0,0 +1,110 @@
+# Mode: batch — Bulk Offer Processing
+
+Two usage modes: **conductor** (navigates portals in real time via Geometra MCP) or **standalone** (script for already-collected URLs).
+
+## Apply The Session-Length Rule
+
+**Never run `batch` as one long interactive session.** Each offer gets its own `opencode run` worker via `batch-runner.sh` — that's the whole point of the architecture. Workers have clean ~200K-token contexts and exit after producing one report + PDF + tracker line, so prompt caching stays healthy.
+
+If you find yourself doing `geometra_fill_form` or `geometra_page_model` for the Nth time in the *same* session, stop and delegate. See "Session Hygiene" in `.opencode/skills/job-forge.md` for the full rationale (cache-bust behavior with repeated Geometra tool calls).
+
+## Use This Architecture
+
+```
+opencode Conductor (opencode --dangerously-skip-permissions)
+  │
+  │  Geometra MCP: navigates portals (logged-in sessions)
+  │  Reads structured page model — the user sees everything in real time
+  │
+  ├─ Offer 1: reads JD from DOM + URL
+  │    └─► opencode run worker → report .md + PDF + tracker-line
+  │
+  ├─ Offer 2: click next, reads JD + URL
+  │    └─► opencode run worker → report .md + PDF + tracker-line
+  │
+  └─ End: merge tracker-additions → data/applications/ + summary
+```
+
+Each worker is a child `opencode run` with a clean 200K token context. The conductor only orchestrates.
+
+## Read These Files
+
+```
+batch/
+  batch-input.tsv               # URLs (from conductor or manual)
+  batch-state.tsv               # Progress (auto-generated, gitignored)
+  batch-runner.sh               # Standalone orchestrator script
+  batch-prompt.md               # Prompt template for workers
+  logs/                         # One log per offer (gitignored)
+  tracker-additions/            # Tracker lines (gitignored)
+```
+
+## Run Mode A Conductor --chrome
+
+1. **Read state**: `batch/batch-state.tsv` → know what has already been processed
+2. **Navigate portal**: Chrome → search URL
+3. **Extract URLs**: Read results DOM → extract URL list → append to `batch-input.tsv`
+4. **For each pending URL**:
+   a. Chrome: click on the offer → read JD text from DOM
+   b. Save JD to `/tmp/batch-jd-{id}.txt`
+   c. Calculate next sequential REPORT_NUM
+d. Execute via Bash:
+       ```bash
+       opencode run --dangerously-skip-permissions \
+         --file batch/batch-prompt.md \
+         "Process this offer. URL: {url}. JD: /tmp/batch-jd-{id}.txt. Report: {num}. ID: {id}"
+       ```
+   e. Update `batch-state.tsv` (completed/failed + score + report_num)
+   f. Log to `logs/{report_num}-{id}.log`
+   g. Chrome: go back → next offer
+5. **Pagination**: If no more offers → click "Next" → repeat
+6. **End**: Merge `tracker-additions/` → `data/applications/` (via `merge-tracker.mjs`) + summary
+
+## Run Mode B Standalone Script
+
+```bash
+batch/batch-runner.sh [OPTIONS]
+```
+
+Options:
+- `--dry-run` — list pending without executing
+- `--retry-failed` — only retry failed ones
+- `--start-from N` — start from ID N
+- `--parallel N` — N workers in parallel
+- `--max-retries N` — attempts per offer (default: 2)
+
+## Read batch-state.tsv Format
+
+```
+id	url	status	started_at	completed_at	report_num	score	error	retries
+1	https://...	completed	2026-...	2026-...	002	4.2	-	0
+2	https://...	failed	2026-...	2026-...	-	-	Error msg	1
+3	https://...	pending	-	-	-	-	-	0
+```
+
+## Use Resumability
+
+- If it dies → re-run → reads `batch-state.tsv` → skips completed
+- Lock file (`batch-runner.pid`) prevents double execution
+- Each worker is independent: failure on offer #47 does not affect the rest
+
+## Run Workers (opencode run)
+
+Each worker receives `batch-prompt.md` as system prompt. It is self-contained.
+
+The worker produces:
+1. Report `.md` in `reports/`
+2. PDF in `output/`
+3. Tracker line in `batch/tracker-additions/{id}.tsv`
+4. JSON result via stdout
+
+## Apply Error Handling
+
+| Error | Recovery |
+|-------|----------|
+| URL inaccessible | Worker fails → conductor marks `failed`, moves on |
+| JD behind login | Conductor tries to read DOM. If it fails → `failed` |
+| Portal changes layout | Conductor reasons about HTML, adapts |
+| Worker crashes | Conductor marks `failed`, moves on. Retry with `--retry-failed` |
+| Conductor dies | Re-run → reads state → skips completed |
+| PDF fails | Report .md is saved. PDF remains pending |

package/modes/compare.md

@@ -0,0 +1,23 @@
+# Mode: compare — Multi-Offer Comparison
+
+**Uses the Canonical Scoring Model from `modes/_shared.md`** — the same 10 weighted dimensions used in `offer`, `auto-pipeline`, and `batch` evaluations. This means scores from prior evaluations are directly comparable. No need to re-score unless the user explicitly requests it.
+
+## Run This Workflow
+
+1. **Gather offers**: Ask the user for offers if not in context. Can be text, URLs, or references to already-evaluated offers in the tracker.
+2. **Score each offer**: If an offer was already evaluated (has a report in `reports/`), read the existing per-dimension scores from the report. If not yet evaluated, score using the full Canonical Scoring Model.
+3. **Build comparison matrix**: Show all 10 dimensions side-by-side across offers, with the weighted total.
+4. **Rank and recommend**: Final ranking with recommendation. Highlight where offers differ most (the dimensions that actually discriminate). Include time-to-offer as a tiebreaker when scores are close (within 0.3).
+
+## Output format
+
+```
+| Dimension (weight) | Offer A | Offer B | Offer C |
+|---------------------|---------|---------|---------|
+| North Star (25%) | X/5 | X/5 | X/5 |
+| CV match (15%) | ... | ... | ... |
+| ... | ... | ... | ... |
+| **Weighted total** | **X.X** | **X.X** | **X.X** |
+```
+
+Then: narrative recommendation explaining which offer to prioritize and why, considering both score and practical factors (timeline, negotiation leverage, start-date flexibility, competing offer stages).

package/modes/contact.md

@@ -0,0 +1,82 @@
+# Mode: contact — LinkedIn Power Move
+
+## Step 0 — Load evaluation context
+
+Check for an existing evaluation report before generating any message.
+
+1. Identify the company + role (from user input, current conversation, or most recent evaluation).
+2. Search `reports/` for a matching report (Grep case-insensitive by company name).
+3. Read any matching report and extract these fields (all block references below point into `modes/offer.md`).
+<!-- isolint-disable-next-line undefined-step-reference -->
+   - **Archetype** detected (Step 0 of evaluation).
+<!-- isolint-disable-next-line undefined-step-reference -->
+   - **Top 3 proof points** from Block B (the JD requirements where CV match was strongest).
+<!-- isolint-disable-next-line undefined-step-reference -->
+   - **Score** and key gaps from Block B.
+<!-- isolint-disable-next-line undefined-step-reference -->
+   - **STAR stories** from Block F that are most relevant.
+<!-- isolint-disable-next-line undefined-step-reference -->
+   - **Case study** recommended in Block F.
+4. Also read all day files in `data/applications/` to check current status of this application.
+5. If NO report exists, inform the user and offer to run an evaluation first — or proceed with cv.md only.
+
+The loaded evaluation context is what makes the outreach message specific instead of generic.
+
+## Step 1 — Identify Targets
+
+Use WebSearch to find:
+
+- Hiring manager of the team.
+- Recruiter assigned to the role.
+- 2-3 peers on the team (people in a similar role).
+
+## Step 2 — Select primary target
+
+Choose the person who would most benefit from the candidate joining. Typically:
+
+- For IC roles: the hiring manager or tech lead.
+- For leadership roles: a peer or the person the role reports to.
+- Avoid cold-messaging the recruiter first unless no other option — a warm intro from a team member is stronger.
+
+## Step 3 — Generate message
+
+Generate the message directly from the evaluation report, not generic claims.
+
+Framework (3 sentences, max 300 characters for LinkedIn connection request):
+
+<!-- isolint-disable-next-line undefined-step-reference -->
+- **Sentence 1 (Hook)**: Something specific about their company or current challenge with AI — NOT generic. If the report's Block A identified the domain and function, reference it.
+<!-- isolint-disable-next-line undefined-step-reference -->
+- **Sentence 2 (Proof)**: The single strongest proof point from Block B's top matches. Use the exact framing that scored highest against the JD. If article-digest.md has a quantified metric for this proof point, use it.
+- **Sentence 3 (Proposal)**: Quick chat, no pressure — "Would love to chat about [specific topic from the JD] for 15 min".
+
+**Archetype-adapted framing** from the report — use the single row matching the detected archetype:
+
+| Archetype | Emphasize |
+|-----------|-----------|
+| FDE | fast delivery, client-facing results |
+| SA | system design, integration wins |
+| PM | product discovery, stakeholder outcomes |
+| LLMOps | production metrics, evals, observability |
+| Agentic | orchestration, reliability, HITL |
+| Transformation | adoption, change management, org impact |
+
+## Step 4 — Generate Versions
+
+Generate:
+- EN (default)
+<!-- isolint-disable-next-line undefined-step-reference -->
+- **Follow-up variant**: A longer version (2-3 sentences) for LinkedIn InMail or email, where the 300-char limit doesn't apply. This version can include a second proof point and a link to the relevant case study from Block F.
+
+## Step 5 — List Alternative Targets
+
+List 2-3 backup contacts with justification for why they're strong second choices.
+
+## Apply These Message Rules
+
+- Max 300 characters for connection request version
+- NO corporate-speak
+- NO "I'm passionate about..."
+- Something that makes them want to respond
+- NEVER share phone number
+- **Every claim must trace back to cv.md or article-digest.md** — no invented metrics

package/modes/deep.md

@@ -0,0 +1,99 @@
+# Mode: deep — Deep Company Research
+
+Structured company research that feeds directly into evaluations and interview prep.
+
+## Step 0 — Check for existing evaluation
+
+Check `reports/` for an existing evaluation of this company+role before researching. If found, load the report — deep research MUST build on what the evaluation already identified (archetype, gaps, proof points), not start from scratch.
+
+## Scan These 6 Research Axes
+
+### Scan Axis 1: AI Strategy
+- What products/features use AI/ML?
+- What's their AI stack? (models, infra, tools)
+- Engineering blog — what do they publish?
+- Papers or talks on AI?
+
+### Scan Axis 2: Recent Moves (last 6 months)
+- Relevant hires in AI/ML/product?
+- Acquisitions or partnerships?
+- Product launches or pivots?
+- Funding rounds or leadership changes?
+
+### Scan Axis 3: Engineering Culture
+- How do they ship? (deploy cadence, CI/CD)
+- Mono-repo or multi-repo?
+- Languages/frameworks?
+- Remote-first or office-first?
+- Glassdoor/Blind reviews on eng culture?
+
+### Scan Axis 4: Likely Challenges
+- Scaling problems?
+- Reliability, cost, latency challenges?
+- Migrating anything? (infra, models, platforms)
+- Pain points from reviews?
+
+### Scan Axis 5: Competitors And Differentiation
+- Main competitors?
+- Moat/differentiator?
+- Positioning vs competition?
+
+### Scan Axis 6: Candidate Angle
+cv.md is already in context via opencode.json:instructions — do NOT Read it again. Read profile.yml once if you need identity fields:
+- What unique value does the candidate bring to this team?
+- Which projects are most relevant?
+- What story should they tell in the interview?
+
+## Feed Research Back Into Evaluation
+
+**The feedback loop is the key step that makes deep research actionable.**
+
+After completing the 6 axes, update the evaluation if one exists:
+
+1. **Score adjustments**: Deep research may reveal information that changes scoring dimensions.
+   - Company reputation (axis 3: culture) → dimension #7.
+   - Tech stack modernity (axis 1: AI strategy) → dimension #8.
+   - Speed to offer (axis 2: recent moves, hiring pace) → dimension #9.
+   - Cultural signals (axis 3: culture) → dimension #10.
+   - Growth trajectory (axis 2: funding, launches) → dimension #5.
+
+2. **Append a `## Deep Research` section to the existing report** with findings organized by axis. Include date of research so it can be refreshed.
+
+<!-- isolint-disable-next-line undefined-step-reference -->
+3. **Update Block F (Interview Plan)**: The "likely challenges" (axis 4) directly inform STAR story selection — pick stories that address the company's actual problems, not generic ones.
+
+4. **Update contact targeting**: The "recent moves" (axis 2) often reveal who to reach out to (new hires, team leads mentioned in blog posts).
+
+<!-- isolint-disable-next-line undefined-step-reference -->
+If NO evaluation exists yet, save the research to `reports/deep-{company-slug}-{date}.md` so it's available when the user evaluates later. The evaluation mode MUST check for existing deep research before starting Block D (Comp & Demand).
+
+## Output Format
+
+```markdown
+## Deep Research: [Company] — [Role]
+**Date:** YYYY-MM-DD
+**Linked report:** #NNN (if exists)
+
+### 1. AI Strategy
+(findings)
+
+### 2. Recent Moves
+(findings)
+
+### 3. Engineering Culture
+(findings)
+
+### 4. Likely Challenges
+(findings)
+
+### 5. Competitors & Differentiation
+(findings)
+
+### 6. Candidate Angle
+(findings)
+
+### Score Impact
+| Dimension | Before | After | Why |
+|-----------|--------|-------|-----|
+(only dimensions where research changed the score)
+```

package/modes/followup.md

@@ -0,0 +1,68 @@
+# Mode: followup — Follow-Up Timing & Nudge System
+
+Scans all day files in `data/applications/` for entries that need follow-up action based on their current state and how long they've been in that state.
+
+**This mode is read-only on existing pipeline logic.** It reads the tracker and suggests actions — it never changes scores, reports, or pipeline behavior.
+
+## Timing Rules
+
+| Current State | Days Since | Action |
+|---------------|-----------|--------|
+| Applied | 7-10 days | Nudge: LinkedIn outreach via `/job-forge contact` if not already Contacted |
+| Applied | 14+ days | Flag as stale. Suggest: nudge or archive to Discarded |
+| Contacted | 5-7 days | Follow-up message (shorter, reference first message) |
+| Contacted | 14+ days | Flag as stale. Likely no response — move on |
+| Responded | 5 days | If no next step scheduled, ask: "Did they propose a call?" |
+| Interview | 1 day after | Send thank-you note (generate draft) |
+| Interview | 7+ days no update | Nudge recruiter: "Following up on our conversation last week" |
+| Evaluated | 14+ days | Stale evaluation. Offer may be closed — suggest verifying or archiving |
+
+## Run This Workflow
+
+1. **Read** all day files in `data/applications/` — parse all entries with dates.
+2. **Compute** days since the date column for each entry (using today's date).
+3. **Filter** to entries matching the Timing Rules table above.
+4. **Sort** by urgency (most overdue first).
+5. **Present** action list.
+
+```
+## Follow-Up Actions — {today's date}
+
+### Urgent (overdue)
+- #045 Anthropic — AI Engineer | Applied 12 days ago → Nudge via LinkedIn
+- #078 Datadog — Staff PM | Interviewed 8 days ago → Follow up with recruiter
+
+### Coming Up (next 3 days)
+- #102 Stripe — AI Platform | Applied 6 days ago → Nudge window opens in 1 day
+
+### Stale (consider archiving)
+- #023 OldCo — Senior Eng | Evaluated 21 days ago → Verify if still open or Discard
+```
+
+## Nudge Message Generation
+
+When the user selects an entry to nudge:
+
+1. Read the existing report from `reports/`.
+2. Use the **contact** mode logic to generate a follow-up message (not a first outreach).
+3. Follow-up messages are shorter and reference the application.
+   - "I applied to [role] [N days] ago and wanted to follow up...".
+<!-- isolint-disable-next-line undefined-step-reference -->
+   - Reference one specific proof point from Block B (see `modes/offer.md`).
+   - Keep it under 200 characters for LinkedIn.
+
+## Generate Thank-You Notes
+
+After interviews, generate a thank-you note.
+<!-- isolint-disable-next-line undefined-step-reference -->
+1. Read the report + Block F STAR stories (Block F lives in `modes/offer.md`).
+2. Reference something specific discussed in the interview (ask the user what stood out).
+3. Reinforce one proof point.
+4. 3-4 sentences max, send within 24 hours.
+
+## Automate This Mode
+
+The followup mode works well with `/loop` or `/schedule`:
+
+- Run `/job-forge followup` every 2-3 days to catch nudge windows.
+- Suggest this to the user after their first batch of applications.

package/modes/negotiation.md

@@ -0,0 +1,146 @@
+# Mode: negotiation — Offer Negotiation
+
+Structured mode for when the user receives an actual job offer. Completely separate from `offer` (which evaluates job postings).
+
+**This mode is standalone.** It reads existing data (reports, tracker, profile.yml) but creates its own output. No changes to existing pipeline behavior.
+
+## Use These Trigger Conditions
+
+- User says "I got an offer from [company]".
+- User updates a tracker entry to Offer.
+- User runs `/job-forge negotiation`.
+
+## Step 1 — Capture The Offer
+
+Ask the user for the offer details (or read from a document/screenshot):
+
+| Field | Example | Notes |
+|-------|---------|-------|
+| Base salary | $180,000 | Annual |
+| Equity/RSUs | $50,000/yr (4yr vest) | Total grant ÷ vesting period |
+| Signing bonus | $20,000 | One-time |
+| Annual bonus | 15% target | As % of base |
+| Benefits | Standard | Healthcare, 401k match, dental, vision |
+| PTO | Unlimited / 20 days | |
+| Remote policy | Full remote | |
+| Start date | YYYY-MM-DD | |
+| Deadline to respond | YYYY-MM-DD | Critical — drives urgency |
+| Level/title | Senior Engineer | |
+| Other perks | $5K learning budget | Anything else |
+
+## Step 2 — Total Comp Calculation
+
+Compute total annual comp:
+
+```
+Base:           $180,000
+Equity (annual): $50,000   (total grant ÷ vesting years)
+Bonus (target):  $27,000   (base × bonus %)
+Signing (annual): $10,000   (signing ÷ 2, amortized over first 2 years)
+───────────────────────────
+Total Year 1:   $267,000
+Total Ongoing:  $257,000   (without signing bonus amortization)
+```
+
+**For equity at private companies:**
+- Apply a discount: pre-IPO equity is illiquid. Common: 50-70% discount for early stage, 20-30% for late stage.
+- Note the strike price if options (not RSUs). Compute current spread if possible.
+- Flag if there's no secondary market.
+
+## Step 3 — Compare Against Target & Pipeline
+
+Read from `config/profile.yml`:
+- `compensation.target_range` — is the offer within, above, or below?
+- `compensation.minimum` — is it above the walk-away number?
+
+Read from all day files in `data/applications/`:
+- Count other offers at Interview or Offer stage (these in-flight offers form your leverage).
+- What scores did those have? (Higher-scored alternatives = stronger negotiating position.)
+
+```
+## Leverage Assessment
+
+Active pipeline:
+- Offer: [this one] + N others
+- Interview stage: N companies
+- Applied (awaiting): N companies
+
+Leverage: [Strong/Moderate/Weak]
+- Strong: 2+ competing offers or 3+ active interviews
+- Moderate: 1 other offer or 2+ interviews
+- Weak: this is the only active opportunity
+```
+
+## Step 4 — Build Counter-Offer Strategy
+
+Build the strategy based on leverage and gap to target.
+
+**If offer is below target range:**
+```
+Recommended counter: [target midpoint + 10%]
+Justification: "Based on market data for [role] at [level], 
+and given [competing offers / pipeline activity], I'm targeting [range]."
+```
+
+**If offer is within target range:**
+```
+Recommended counter: [target top end]
+Justification: "I'm excited about this role. To make the decision 
+straightforward, [specific ask — e.g., bump equity, signing bonus, level]."
+```
+
+**If offer is above target range:**
+```
+Focus negotiation on non-comp factors: level/title, scope, 
+team choice, start date flexibility, remote policy.
+```
+
+### Generate Negotiation Scripts (adapted from _shared.md)
+
+Generate 3 scripts ready for use:
+
+1. **Initial response** (buy time): "Thank you for the offer. I'm very interested and want to give this proper consideration. Can I have until [deadline or +3 days] to review the full package?"
+
+<!-- isolint-disable-next-line undefined-step-reference -->
+2. **Counter-offer** when comp is below the target range: Specific to the gap. Reference market data from Block D (see `modes/offer.md`) when available. Never give a single number — always a range.
+
+3. **Acceptance** (when ready): Confirm all terms in writing. Ask for the offer letter with the agreed terms before giving verbal acceptance.
+
+### Skip These Negotiation Anti-Patterns
+- Never counter before understanding the full package
+- Never reveal your current comp or other offer amounts
+- Never give an ultimatum unless you're prepared to walk
+- Never negotiate via email if you can do it via call (nuance matters)
+
+## Step 5 — Build Decision Framework
+
+Build a decision matrix when the user has multiple offers, using the same scoring dimensions from `_shared.md`.
+
+```
+| Dimension | Offer A | Offer B |
+|-----------|---------|---------|
+| Total comp (Year 1) | $267K | $245K |
+| Equity upside | Moderate (Series C) | High (Series A) |
+| Role match (from eval) | 4.5/5 | 3.8/5 |
+| Growth trajectory | Clear path to Staff | Flat org |
+| Remote quality | Full remote | Hybrid 2x/week |
+| Gut feeling | High | Medium |
+```
+
+Include a "gut feeling" row — quantitative analysis doesn't capture everything.
+
+## Step 6 — Save And Update Tracker
+
+Update the tracker entry to reflect the offer.
+
+1. Update the entry in its day file under `data/applications/YYYY-MM-DD.md`: status → `Offer`, notes → comp summary.
+2. Append `## Offer Details` section to the existing report with full comp breakdown and negotiation outcome.
+3. If the user accepts: status → a new note like `Accepted YYYY-MM-DD`.
+
+## Manage The Deadline
+
+When the user has a response deadline, run these checks in order.
+
+1. Calculate business days remaining.
+2. If < 3 business days remain AND other interviews are active, suggest: "Ask [company] for an extension — 'I want to make a thoughtful decision and have a final interview scheduled this week. Could we extend to [date]?'"
+3. If the deadline is firm and leverage is weak, help the user decide with what they have.