cost-estimate（成本估算）

Non-negotiable rules:

1. Use the helper scripts under .agents/skills/cost-estimate/helpers/ instead of redoing LOC, session, or cost math manually.
2. Keep scope explicit: full repo, branch:<name>, or commit:<hash>.
3. Separate raw engineering hours from organizational overhead and team multipliers.
4. Keep the pricing rubric and report schema in references, not inline in the invocation path.
5. Treat external market-rate research as optional. Use built-in rates unless the user explicitly requests a different market or region.

Cost Estimate

Inputs

• $request: Optional scope or estimation guidance such as branch:feat/foo, commit:abc1234, region hints, or desired audience

Goal

Produce a credible estimate that:

• measures the requested code scope with the helper scripts
• classifies code into the right productivity buckets
• computes engineering hours and costs without double-counting overhead
• translates that estimate into realistic calendar and team-cost views
• reports assumptions, confidence, and Claude ROI clearly

Step 0: Resolve scope and mode

Parse the request into one of:

• full repository
• branch:<name>
• commit:<hash>

If the request is ambiguous, infer full repository by default. If the user supplied region or market hints, carry them into the final rate discussion. Otherwise use the built-in baseline rates.

Success criteria: The estimation target and pricing basis are explicit before running scripts.

Step 1: Measure the real code surface with the helper scripts

Use the helper toolchain in .agents/skills/cost-estimate/helpers/:

• loc_counter.py
• git_session_analyzer.py
• cost_calculator.py
• report_generator.py

Run loc_counter.py for the resolved scope to capture:

• total lines
• file counts
• language breakdown
• directory breakdown
• source vs test vs config vs docs
• per-file category candidates from all_files

Run git_session_analyzer.py when git history is available so Claude ROI can use actual session estimates rather than LOC fallback.

Rules:

• prefer the script output over ad hoc find, wc, or inline arithmetic
• if scope is a branch or commit, measure the diff rather than the full repository
• if git history is missing or unhelpful, fall back to LOC-based Claude hour estimates later

Success criteria: You have structured JSON for scope size and, when possible, active-session estimates.

Step 2: Classify the code into productivity buckets

Review the loc_counter.py output and map every relevant source line into exactly one category.

Use references/estimation-rates.md for:

• category keys
• productivity ranges
• overhead ranges
• market-rate baselines
• role multipliers
• organizational efficiency constants
• Claude ROI constants

Classification rules:

• assign each source line once
• keep tests, config/build, and documentation separate from product code
• detect specialized work such as GPU, native interop, audio/video, system extensions, or on-device ML
• do not inflate complexity without file-level evidence
• group files when a directory clearly shares one category, but call out exceptional files separately

Success criteria: There is a complete category-to-line-count JSON payload for the calculator.

Step 3: Run the calculator and sanity-check the output

Pipe the category totals into cost_calculator.py.

The calculator should produce:

• base coding hours
• overhead hours
• total estimated hours
• sanity-check effective lines/hour
• calendar-time tables
• engineering-only cost
• full-team cost
• Claude ROI fields when Claude hours are available

Rules:

• do not bake overhead into the category assignment; the calculator already adds overhead
• if the sanity check falls outside the target range, adjust category assignments or explain why this repo is legitimately outside the norm
• use built-in market rates by default unless the user explicitly requested a different market basis

Success criteria: The calculator output is internally consistent and the sanity check has been reviewed.

Step 4: Refine Claude ROI and confidence

If session data exists, inspect it and adjust only when the default commit-density heuristic is obviously understating large-scope work. If session data is missing, use the fallback Claude productivity constant from references/estimation-rates.md.

Report:

• estimated Claude active hours
• speed multiplier vs the baseline human rate
• value per Claude hour
• headline ROI and savings

Also state confidence:

• high when scope, category mix, and history are clean
• medium when history or categorization is incomplete
• low when the request is intentionally approximate or the repo is only partially available

Success criteria: The ROI story is explicit, bounded, and not overstated.

Step 5: Generate the report body

Use report_generator.py to generate the markdown backbone, then refine the narrative where needed.

Load references/report-contract.md for the required report structure and minimum sections.

The final estimate should cover:

• executive summary
• codebase metrics
• development-time estimate
• calendar-time view
• engineering-only cost
• full-team cost
• Claude ROI
• assumptions and caveats

Rules:

• lead with the executive summary and Claude ROI
• keep the report stakeholder-readable, not tool-dump heavy
• mention the scope basis explicitly
• preserve escaped currency formatting when editing prose manually

Success criteria: The estimate is readable, structured, and aligned with the report contract.

Guardrails

• Do not add disable-model-invocation; this is a read-heavy analysis workflow.
• Do not add paths:; this is a generic estimation skill.
• Do not keep pricing tables, role matrices, or full report templates inline in SKILL.md.
• Do not replace helper-script output with manual math unless the helper chain is unavailable.
• Do not present a cost number without the scope, assumptions, and confidence level.
• Do not claim region-specific market validation unless the user explicitly requested it and that research was actually performed.

When To Load References

• references/estimation-rates.md Use for the productivity buckets, overhead rates, market-rate baselines, team multipliers, efficiency constants, and Claude ROI fallback constants.
• references/report-contract.md Use for the mandatory section order, reporting contract, and required caveats.

Helper Scripts

• .agents/skills/cost-estimate/helpers/loc_counter.py

- full repo: no flags - branch diff: --branch <name> (optionally --base <base>) - single commit: --commit <hash> - output: JSON with totals, by_language, by_directory, all_files

• .agents/skills/cost-estimate/helpers/git_session_analyzer.py

- all commits: no flags - specific branch: --branch <name> - output: JSON with total_commits, total_sessions, estimated_active_hours, sessions[]

• .agents/skills/cost-estimate/helpers/cost_calculator.py

- input: pipe category JSON on stdin - flags: --rate <hourly>, --claude-hours <N> - valid category keys: simple_crud_ui_boilerplate, standard_views, complex_ui, business_logic, database_persistence, audio_video_processing, gpu_shader, native_interop, system_extensions, on_device_ml, tests, config_build, documentation - output: JSON with base_coding, overhead, total_estimated_hours, sanity_check, calendar_time, engineering_cost, team_costs, claude_roi

• .agents/skills/cost-estimate/helpers/report_generator.py

- flags: --calc <costs.json>, --sessions <sessions.json>, --project <name>, --scope <desc> - single section: --section <name> - available sections: executive_summary, development_time, calendar_time, engineering_cost, team_cost, grand_total, claude_roi, assumptions - output: ready-to-paste markdown

Use these directly. The judgment work in this skill is classification, calibration, and explanation, not reimplementing the scripts.

Output Contract

Report:

1. resolved scope and pricing basis
2. key repo metrics and complexity drivers
3. engineering hours and sanity-check result
4. engineering-only and team-cost ranges
5. Claude ROI and confidence level
6. assumptions, caveats, and any missing-data limitations