tokencast 0.1.0

Pre-execution cost estimation for LLM agent workflows with calibration learning

tokencast

A Claude Code skill that estimates Anthropic API cost for planned agent tasks, then learns from actual usage to improve estimates over time.

Install once per project. It auto-estimates after plans are created and auto-learns at session end. Zero ongoing friction.

Setup (one time per project)

# Clone the repo (anywhere — it doesn't need to live inside your project)
git clone https://github.com/krulewis/tokencast.git

# Install into your project (quote paths with spaces)
bash tokencast/scripts/install-hooks.sh "/path/to/your-project"

Paths with spaces: Always wrap the project path in quotes. Without them the install script will fail on paths like /Volumes/Macintosh HD2/....

This does three things:

1. Symlinks the skill into <project>/.claude/skills/tokencast/
2. Adds a Stop hook for auto-learning at session end
3. Adds a PostToolUse hook to nudge estimation after planning agents

Every Claude Code session in that project now has tokencast active.

What Happens Automatically

After a plan is created

tokencast detects the plan in conversation context, infers size, files, complexity, project type, and language, then outputs a cost table:

## tokencast estimate

Change: size=M, files=5, complexity=medium
Calibration: 1.12x from 8 prior runs

| Step                  | Model  | Optimistic | Expected | Pessimistic |
|-----------------------|--------|------------|----------|-------------|
| Research Agent        | Sonnet | $0.60      | $1.17    | $4.47       |
| Architect Agent       | Opus   | $0.67      | $1.18    | $3.97       |
| ...                   | ...    | ...        | ...      | ...         |
| TOTAL                 |        | $3.37      | $6.26    | $22.64      |

At session end

The learning hook silently:

1. Reads the session's JSONL log
2. Computes actual token cost (including cache write tokens)
3. Compares to the estimate
4. Updates calibration factors

Next session

Future estimates use learned correction factors. More sessions = better accuracy.

Manual Invocation

You can also invoke explicitly with overrides:

/tokencast size=L files=12 complexity=high
/tokencast steps=implement,test,qa
/tokencast review_cycles=3
/tokencast review_cycles=0

Use review_cycles=N to set the number of expected PR review cycles. Use review_cycles=0 to suppress the PR Review Loop row.

How It Works

1. Infers size, file count, complexity from the plan in conversation
2. Reads reference files for pricing and token heuristics
3. Loads learned calibration factors (if any exist)
4. Computes per-step token estimates using activity decomposition
5. Applies complexity multiplier, context accumulation (K+1)/2, and cache rates
6. Splits into Optimistic / Expected / Pessimistic bands
7. If PR Review Loop is in scope, computes loop cost using geometric decay across N review cycles (Optimistic=1, Expected=N, Pessimistic=N×2)
8. Applies calibration correction to Expected band (individual steps re-anchor; PR Review Loop scales each band independently)
9. Records the estimate for later comparison with actuals

Overrides

Override	Effect
size=M	Set size class explicitly
files=5	Set file count explicitly
complexity=high	Set complexity explicitly
steps=implement,test,qa	Estimate only those pipeline steps
project_type=migration	Set project type explicitly
language=go	Set primary language explicitly
review_cycles=3	Set PR review cycle count (0 = disable)

Confidence Bands

Band	Cache Hit	Multiplier	Meaning
Optimistic	60%	0.6x	Best case — focused agent work
Expected	50%	1.0x	Typical run
Pessimistic	30%	3.0x	With rework loops, debugging, retries

Calibration

Calibration is fully automatic:

• 0-2 sessions: No correction applied. "Collecting data" status.
• 3-10 sessions: Global correction factor via trimmed mean of actual/expected ratios (trim_fraction=0.1).
• 10+ sessions: EWMA with recency weighting. Per-size-class factors activate when a class has 3+ samples.
• Outlier filtering: Sessions with actual/expected ratio >3.0x or <0.2x are excluded from calibration and logged for inspection.

Calibration data lives in calibration/ (gitignored, local to each user).

Disabling

bash /path/to/tokencast/scripts/disable.sh /path/to/your-project

Removes the skill and hooks. Preserves calibration data for reuse.

Files

SKILL.md                        — Skill definition (auto-trigger, algorithm)
references/pricing.md           — Model prices, cache rates, step→model map
references/heuristics.md        — Token budgets, pipeline decompositions, multipliers
references/examples.md          — Worked examples with arithmetic
references/calibration-algorithm.md — Detailed calibration algorithm reference
commands/
  tokencast-version.md     — /tokencast-version slash command
scripts/
  install-hooks.sh              — One-time project setup
  disable.sh                    — Remove from project
  tokencast-learn.sh       — Stop hook: auto-captures actuals
  tokencast-track.sh       — PostToolUse hook: nudges estimation after plans
  sum-session-tokens.py         — Parses session JSONL for actual costs
  update-factors.py             — Computes calibration factors from history
calibration/                    — Per-user local data (gitignored)
  history.jsonl                 — Estimate vs actual records
  factors.json                  — Learned correction factors
  active-estimate.json          — Transient marker for current estimate

v1.1 Changes

• Trimmed mean replaces median for faster convergence with small samples
• Outlier flagging — extreme ratios (>3.0x or <0.2x) excluded from calibration, logged for inspection
• Richer data — project type, language, pipeline signature, and step count captured per session
• Baseline subtraction — tokens spent before the estimate are excluded from actuals
• Security hardening — path injection fixes, consolidated parsing, safe handling of paths with spaces
• Version markers — version: 1.1.0 in SKILL.md, --version flag on learn script

v1.2 Changes

• PR Review Loop modeling — geometric-decay cost model for review-fix-re-review cycles
• New override — review_cycles=N to set expected cycle count (0 = disable)
• Per-band calibration — PR Review Loop applies calibration independently per band (not re-anchored)
• New schema fields — review_cycles_estimated and review_cycles_actual in active-estimate.json

Limitations

• Pipeline step names reflect a default workflow — map your own steps to the closest defaults. Formulas are pipeline-agnostic (see references/heuristics.md)
• Heuristics assume typical 150-300 line source files
• Does not model parallel agent execution
• Calibration requires 3+ completed sessions before corrections activate
• Pricing data embedded; check last_updated in references/pricing.md
• Multi-session tasks only capture the session containing the estimate

License

MIT