| name | planner-management |
| description | How to introspect and control the forecaster, manage endeavour lifecycle, adjust feature weights, edit tasks, view forecast breakdowns, and handle failures. |
| agents | project_manager, cognitive |
Understanding the Forecaster
The Forecaster evaluates active tasks and endeavours using D' scoring with
configurable feature sets. Five default features:
| Feature | Default | Critical | What it measures |
|---|
| step_completion_rate | HIGH | Yes | Steps vs cycles — are you making progress? |
| dependency_health | HIGH | Yes | Blocked dependencies or active blockers |
| complexity_drift | MEDIUM | No | Actual cycles vs planned complexity |
| risk_materialization | MEDIUM | No | How many predicted risks actually happened |
| scope_creep | LOW | No | Steps/phases added beyond the original plan |
D' score = importance-weighted magnitude sum, normalised to [0,1].
Score above replan_threshold (default 0.55) triggers a replan suggestion.
Inspecting Forecaster Config
Use get_forecaster_config to see the global default features and thresholds.
Pass endeavour_id to see the effective config for a specific endeavour
(base defaults merged with per-endeavour overrides).
Viewing Forecast Breakdowns
Use get_forecast_breakdown(id) to see per-feature detail for any task or
endeavour. The breakdown shows each feature's raw magnitude, weighted score
(normalised contribution to the composite D' score), and rationale. This
tells you exactly which feature is driving the health score and why.
Example: if step_completion_rate has magnitude=7 and weighted=0.47 but
scope_creep has magnitude=1 and weighted=0.02, the problem is slow
progress, not scope changes — adjust your response accordingly.
Adjusting Feature Weights
Per-endeavour overrides let you tune the Forecaster without changing global
defaults. Use update_forecaster_config on a specific endeavour.
When the Forecaster is too sensitive (frequent false replan suggestions):
- Increase the threshold:
update_forecaster_config(endeavour_id, threshold_override: 0.70)
- Downgrade noisy features:
update_forecaster_config(endeavour_id, feature_name: "scope_creep", importance: "low")
When the Forecaster misses real problems:
- Lower the threshold:
update_forecaster_config(endeavour_id, threshold_override: 0.40)
- Upgrade critical features:
update_forecaster_config(endeavour_id, feature_name: "dependency_health", importance: "high")
Example: For an endeavour where scope changes are expected (exploratory
research), reduce scope_creep importance to LOW so it doesn't trigger replans.
Task Management
Edit tasks: update_task(task_id, title?, description?) changes title or
description. add_task_step(task_id, description) adds a new step.
remove_task_step(task_id, step_index) removes a step by index.
Quick operations (on cognitive loop, no delegation needed):
complete_task_step — mark a step done
activate_task — set as current focus
get_active_work — list all active tasks and endeavours
get_task_detail — full task structure
Heavy operations (require delegation to project_manager):
create_endeavour — new multi-task initiative
flag_risk — record a materialised risk
abandon_task — stop tracking
request_forecast_review — check plan health scores
get_forecast_breakdown — per-feature D' breakdown for a task or endeavour
delete_task — permanently remove a task (prefer abandon_task for normal cancellation)
delete_endeavour — permanently remove an endeavour (does not delete associated tasks)
Endeavour Lifecycle
Creating: create_endeavour for multi-task initiatives. Then add_phase
to structure the work into ordered phases.
Updating: update_endeavour(endeavour_id, goal?, deadline?, update_cadence?)
to adjust goal, deadline, or stakeholder update frequency after creation.
Scheduling: schedule_work_session for autonomous sessions at specific
times. cancel_work_session to remove a scheduled session. list_work_sessions
to see session history.
Monitoring: get_endeavour_detail for full state (phases, blockers,
sessions, metrics). get_forecaster_config(endeavour_id) for health scoring
config.
Phase Management
add_phase — add a new phase to an endeavour
advance_phase — mark current phase complete, advance to next
(may require operator approval if phase_transition = require_approval)
Handling Failures
Blockers: Use report_blocker to record what's blocking progress. Set
requires_human: true if operator intervention is needed. Use resolve_blocker
when the issue is resolved — this auto-restores Active status if all blockers
are clear.
Session failures: Recorded automatically. Review with list_work_sessions.
Consider cancelling remaining scheduled sessions if the approach needs to change.
Replan triggers: When the Forecaster suggests replanning:
- Check the per-feature breakdown:
get_forecast_breakdown(endeavour_id)
- Identify which features are driving the score (high weighted_score)
- Decide: adjust thresholds (if the signal is noise) or replan (if genuine)
- If replanning: create new phases via
add_phase, adjust the endeavour
Sprint Contracts
When delegating multi-step work to the project_manager, the PM will state
its intent before executing: what it will do, what success looks like, and
what it's assuming. This is the sprint contract pattern.
If the PM's stated plan looks wrong, cancel it (cancel_agent) and
re-delegate with clearer instructions. If it looks right, let it execute.
When you delegate complex work yourself (not through PM), apply the same
principle: state what you're about to do, then do it. Don't silently execute
a long sequence of operations.
Cost Awareness
Each work session consumes cycles and tokens (visible in list_work_sessions
and get_endeavour_detail). Set max_cycles per session to control cost.
Consider cancelling low-value scheduled sessions when budget is tight.
Check the sensorium's tokens_remaining before scheduling heavy sessions.
Task vs Endeavour Decision
- Single sequential plan, completable in one session = Task
- Multiple independent work streams toward one goal = Endeavour
- Research spanning multiple sessions = Endeavour
- One-off investigation = Task
- Deadline-driven multi-phase project = Endeavour