| name | haipipe-end-post |
| description | PostFn specialist -- designs/reviews the response-formatting function in an Endpoint_Set. One of 5 inference Fn-types. Called by /haipipe-end when intent references PostFn, response formatting, post-processing, or `post`. |
| argument-hint | [verb] [use_case] [args...] |
| allowed-tools | Bash, Read, Write, Edit, Grep, Glob |
| metadata | {"version":"0.1.1","last_updated":"2026-07-08","summary":"PostFn specialist — design and review of the response-formatting function in an Endpoint_Set."} |
Skill: haipipe-end-post
Per-Fn-type specialist for PostFn — the inference function that formats the model's raw prediction into the wire response shape returned to the caller.
One of the 5 inference Fn-types inside an Endpoint_Set.
See ref/concepts.md for PostFn semantics.
Verb axis: design | review | list | concepts
Use case: each PostFn impl is scoped to ONE response shape (CGM forecast, optimal-
message bandit / holistic, R3 SMS lineups, weight-loss multi-label,
xgboost multi-models). design and review take a use_case argument.
Commands
/haipipe-end-post -> show PostFn ref (concepts mode)
/haipipe-end-post concepts -> same
/haipipe-end-post list -> list known use-case impls in fn_post/
/haipipe-end-post design <use_case> [endpoint_set] -> scaffold a new PostFn for <use_case>
/haipipe-end-post review <use_case> [endpoint_set] -> structural audit of one use-case impl
Use cases (concrete impls in code/haifn/fn_endpoint/fn_post/, as of 2026-04-25 — discover current impls with ls code/haifn/fn_endpoint/)
CGMForecast_v260101 CGM forecast response
OptimalMessage_Bandit_v250620 bandit-driven optimal message
OptimalMessage_Bandit_Greedy_v250620 bandit (greedy variant)
OptimalMessage_Bandit_ABCTest_v250620 bandit (ABC test variant)
OptimalMessage_Holistic_Greedy_v250721 holistic (greedy variant)
OptimalMessage_Holistic_ABTest_v250721 holistic (AB test variant)
OptimalMessage_Holistic_ABCTest_v250721 holistic (ABC test variant)
R3sms_9o20_ArmGreedy_v250922 R3 SMS 9-of-20 greedy
R3sms_9o20_ArmGreedyRandom_v250922 R3 SMS 9-of-20 greedy + random
R3sms_9o20_ABTest_v250922 R3 SMS 9-of-20 AB test
R4sms_10o40_ArmGreedy_v260807 R4 SMS 10-of-40 greedy (top-9 + default)
WeightLossMultiLabel_PostFn_v260305 weight-loss multi-label
WeightLossMultiLabel_Af1M_PostFn_v260310 weight-loss (Af1M variant)
WeightLossMultiLabel_v3_PostFn_v260316 weight-loss v3
XgboostMultiModels_PostFn_v0610 xgboost multi-model ensemble
If <use_case> is omitted, the skill should Bash("ls code/haifn/fn_endpoint/fn_post/") and ask the user to pick.
Dispatch Table
Verb Reads
---------- ------------------------------------------------------------------
design ref/concepts.md
../haipipe-end/fn/fn-design.md
../haipipe-end/ref/0-overview.md
review ref/concepts.md
../haipipe-end-endpointset/fn/fn-review.md
concepts ref/concepts.md (only)
Step-by-Step Protocol
Step 0: Read ref/concepts.md — PostFn semantics, response schema, formatting rules.
Step 1: For design, also read ../haipipe-end/fn/fn-design.md + ../haipipe-end/ref/0-overview.md.
For review, also read ../haipipe-end-endpointset/fn/fn-review.md.
Step 2: Execute the procedure scoped to PostFn.
Step 3: Emit the structured tail.
Guardrails (learned the hard way — do NOT skip)
POST-1 An ARM-RESTRICTING PostFn must RAISE on a candidate that does not resolve.
The `action_list_candidates` pattern filters a hardcoded name list by
membership in the model's arm set:
action_list = [i for i in action_list_candidates if i in action_list_full]
A name that is absent is dropped SILENTLY. When SMSR4 expanded 20 arms ->
40 and renamed one, a 3-name list became a 2-name list and the endpoint
served `salience` on 6,513/6,513 live requests for months. The list was
never wrong for the round it was written for; it was wrong for the round
it was INHERITED into, and nothing in the code could tell the difference.
Always assert first:
_unresolved = [i for i in action_list_candidates if i not in action_list_full]
assert not _unresolved, f'candidate arms absent from model output: {_unresolved}'
POST-2 One PostFn per (round, candidate set, policy). NEVER edit a shared impl
in place. Count the manifests first:
find . -name manifest.json | xargs grep -l '"PostFn": "<name>"'
Nine manifests shared OptimalMessage_Holistic_Greedy_v250721, and its
list was still CORRECT for SMSR3. Editing it to fix SMSR4 would have
broken SMSR3. Fork with a dated name following the existing convention:
`<Round>_<N>o<M>_<Policy>_v<YYMMDD>` (R3sms_9o20_ArmGreedy_v250922).
POST-3 A restricted candidate list constrains ONLY `action.name`. The response
still carries every arm in `predictions[]`. So a broken candidate list
leaves all scores correct and serves the wrong message: any check that
compares only the returned SCORES cannot see it. Assert on the SERVED arm.
(See haipipe-task guardrail GATE-1.)
POST-4 Derive the candidate list from an eval artifact at build time and assert
it against a frozen expected list, rather than typing names in. A hand-
typed list is what goes stale; a derived one regenerates and a drifted
eval table then FAILS the build instead of shipping quietly.
Scope
Owns:
- PostFn concept ref (
ref/concepts.md)
- PostFn design + review scoped to ONE Fn-type
Does NOT own:
- Other 4 Fn-types — sibling skills
-meta, -trig, -src2input, -input2src
- Whole-artifact verbs —
/haipipe-end-endpointset
- Deployment —
/haipipe-end-deploy-*
Builder templates (code/scripts/haibuilder/)
Copy-and-customize starting point for new PostFn builders:
code/scripts/haibuilder/6-endpoint/c1_build_postfn_weight_multilabel.py ← WellDoc weight PostFn
Project-specific builders live in the task folder:
examples/<project>/tasks/C01_*/00_endpoint_set_fn_develop/c1_build_postfn.py
